| Journal of Smart Sensors and Computing
Received: 20 November 2025; Revised: 29 December 2025; Accepted: 30 December 2025; Published Online: 30 December 2025.
J. Smart Sens. Comput., 2025, 1(3), 25207 | Volume 1 Issue 3 (December 2025) | DOI: https://doi.org/10.64189/ssc.25213
© The Author(s) 2025
This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0)
Multi-Class Skin Lesion Classification Using Transfer
Learning with EfficientNet-B3 and Convolutional
Block Attention Module
Sneha R. Shegar and Supriya S. Patil*
Department of Computer Engineering, Samarth College of Engineering & Management, Pune, Maharashtra, 412410, India
*Email: profsupriyapatil@gmail.com (Supriya S. Patil)
Abstract
Skin diseases represent a significant global health challenge; however, precise automated detection of cutaneous
lesions remains difficult due to high intra-class variability, inter-class similarity, and severe class imbalance
across disease categories. This paper presents a multi-class skin lesion classification framework based on
transfer learning, which integrates an EfficientNet-B3 backbone with a Convolutional Block Attention Module
(CBAM) to enhance the learning of discriminative features. EfficientNet-B3, pre-trained on large-scale natural
image datasets, serves as a powerful feature extractor, while CBAM improves feature representation by
adaptively emphasizing informative channels and spatial locations. This enables the network to focus on
diagnostically relevant lesion regions while suppressing background artifacts. The proposed model is trained
and evaluated on the DermNet-23 dataset, comprising 23 clinically significant skin disease classes. To address
the challenges of multi-class classification and class imbalance, performance is assessed using standard metrics
including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC).
Experimental results demonstrate that the EfficientNet-B3 + CBAM model achieves 87.1% accuracy, 85.6%
macro-F1 score, and 0.94 AUC, outperforming baseline CNN, ResNet50, MobileNetV3, and standard EfficientNet-
B3 models. These results highlight the effectiveness of attention-guided transfer learning for developing robust
and scalable computer-aided diagnostic systems for skin lesion classification.
Keywords: Skin lesion classification; EfficientNet-B3; CBAM; Transfer learning; Computer-aided diagnosis.
1. Introduction
1.1 Clinical background and dermoscopy
Dermoscopy is an imaging modality which is not invasive and which enlarges and improves visualization of the
subcutaneous skin structures in order to allow a more detailed evaluation of pigment patterns, vascular
organization, and lesion boundaries.
[5]
Dermoscopy has a significant higher diagnostic sensitivity and specificity
of melanoma and other pigmented lesions in the hands of experts than the unaided eye view.
[6]
Nonetheless,
dermoscopy interpretation is very operator-specific, and it involves a lot of training and experience. Even
trained dermatologists show inter-observer variability because of minor and overlapping morphological
patterns in benign and malignant lesions. Furthermore, differences in imaging equipment, conditions of
acquisition and the type of skin of the patient are other complicating factors which make standardized visual
assessment impossible.
[7]
Clinicians in primary care and in a resource-constrained setting might not receive the benefit of such higher
levels of dermoscopic training, resulting in either under-referral of suspicious lesions or over-referral of benign
lesions, which has both clinical and economic implications. The growing rate of skin lesion images that have
been obtained through dermatoscopes and consumer-grade cameras increase the necessity of computation
support that is scaled. It, therefore, follows that there is an enthusiastic impetus to build automated
dermoscopic analysis frameworks capable of estimating and/or adding to masterful performance and help to
standardize and equalize diagnostic procedures.
1.2 Challenges in manual dermoscopic assessment
Manual dermoscopic analysis is objective and it is subject to cognitive bias including anchoring, fatigue and
heuristic reliance. The difference between early melanoma and benign nevi or between inflammatory
dermatoses and infectious or neoplastic lesions is often based on subtle textural, chromatic and structural
effects that are not always easily identified. There is also high intra-class variability (e.g., different appearances
of melanoma in different sites on the body and in different skin tones) and inter-class similarity (e.g., benign
lesions that resemble malignancy) that further contribute to obstacles to proper visual diagnosis.
[8]
Additionally,
non-dermoscopic images do not invariably correspond to the diagnostic criteria that are based on dermoscopy
e.g. pattern analysis, algorithmic scoring systems, or the ABCD rule, and implementing these diagnostic criteria
as a system in a high-volume clinical setting can be challenging. With the increase in image repositories, manual
inspection and triage are no longer feasible, and computer-aided diagnosis (CAD) systems that can process
large-scale image streams and maintain the same level of performance are sought. All these problems highlight
the importance of having powerful, data-driven procedures capable of training discriminative patterns to go
beyond handcrafted specifications.
[9]
1.3 Limitations of existing machine learning approaches
Original computational techniques to analyse skin lesions used conventional machine learning pipelines of
hand-engineered feature-extractors (e.g., colour histograms, texture descriptors, border irregularity measures)
and classifier (e.g. support vector machines, k-nearest neighbours or random forests).
[10-14]
Although these
methods have given a first-time understanding of whether automated lesion recognition is feasible, their
performance was inherently limited by the expressiveness of manually specified features.
[15]
Hand-crafted
descriptors typically do not encode high-order and complicated interaction between local patterns, and are
vulnerable to changes in illumination, scale, and camera. Moreover, in classical pipelines, the explicit
segmentation of the lesion area against the skin around it is performed, which is also a non-trivial and non-
error-prone step on its own. These errors in segmentation spread to lower-level features and worsen
classification. Class imbalance also poses a challenge to many traditional algorithms when working with real-
world skin lesion data since the number of benign lesions is many times the number of cases of malignant
lesions. Consequently, these techniques are generally insensitive to rarity of occurrence but clinically important
classes, and are not sufficiently robust to be used in heterogeneous clinical environments.
[16]
1.4 Limitations of existing deep learning approaches
Deep learning and convolutional neural networks (CNNs), in particular, have significantly improved the medical
image analysis and dermoscopic lesion classification performance.
[17,18]
Traditional architectures like VGG,
ResNet, and Inception are highly accurate on hand-curated benchmark data.
[19]
Nonetheless, a number of
constraints exist when such models are implemented on realistic multi-class skin disease data that has a large
range of dermatological diseases. First, most of the previous studies are dealing with binary or low-cardinality
classification (e.g., melanoma vs. benign),
[20,21]
which are not representative of the entire range of dermatological
diagnoses seen in practice. The generalization of such models to multi-class problems in which dozens of types
of diseases are involved creates serious difficulties in learning the discriminatory features of visually similar
classes. Second, vanilla CNN architectures commonly assume that all spatial locations and feature channels are
equal, and there is no explicit representation of the relative significance of various regions and modalities of the
lesion image. The inability results in inefficient use of both local and global contextual information especially in
the presence of background artifacts (hair, rulers, markers or normal skin structures). Third, generic image
classification tasks are traditionally based on conventional deep networks, which are not optimized on a
systematic basis on depth, width, and resolution to the particular limitations of dermatological data. The risk
associated with over-parameterized model is that it overfits small or moderate size clinical data, whereas the
risk associated with under-parameterized architecture is that the model does not have enough capacity to learn
the more complex patterns of lesions. Moreover, most of the existing methods fail to solve specific domain-
related problems which include extreme imbalance of classes, uneven image quality and presence of non-
clinical artifacts within large repositories of skin images.
[22]
1.5 Need for robust automated multi-class classification
The above shortcomings encourage the emergence of effective automated classification schemes to suit the case
of multi-class skin disease detection in diverse image data. An effective CAD system in the field of dermatology
must meet a range of criteria: (1) high discriminative accuracy on a broad range of lesion types, including rare
but high-risk cases; (2) robustness to noise, changing illumination and acquisition artifacts; (3) effective
exploitation of labelled data through transfer learning with large natural image corpora; and (4) architectural
features to concentrate computational resources on areas of lesion of diagnostic interest as opposed to
unproductive background. In addition, these systems should be measured based on clinical priorities, such as
malignant and severe inflammatory or infectious disease sensitivity, and macro-averaged metrics, which
considers class imbalance. In multi-class scenarios, where further datasets such as DermNet-23, with over
twenty differing diagnostic cases, the heterogeneous visualization of the data has to be managed, yet not
favoring the majority classes. Such demands demand the high-level network designs involving the combination
of parameter-efficient backbones and explicit attention mechanisms as well as the strict optimization
techniques.
[23]
1.6 Problem statement
Nonetheless, even with much advancement in deep learning-based skin lesion analysis, some critical gaps are
found in research. To begin with, a comparative lack of approaches incorporating modern and compound-scale
architectures, including EfficientNet-B3,
[24]
alongside attention mechanisms with attention to dermatological
image properties, is relatively scarce. Most of the extant literature is based on the application of legacy CNN
backbones or the application of attention modules narrowly or in an ad-hoc fashion, without a thorough
examination of their effect on multi-class performance. Second, previous research tends to focus on the general
accuracy on a small sample of lesion types, without giving much indication of per-class accuracy and the model
behavior under strong imbalance between classes. Detailed studies based on macro-averaged precision, recall,
F1-score, and area under the ROC curve are required, especially when data includes different and uneven
diagnostic categories. Third, not many studies offer an end-to-end, reproducible pipeline to include effective
preprocessing, systematic data augmentation, optimized transfer learning, and attention-focused feature
refinement to large multi-class dermatology datasets. Lastly, it is not well-known how lightweight and attention-
enhanced architectures can help fill the gap between research prototypes and clinically viable CAD systems.
Namely, how the integration of parameter-efficient backbones and channelspatial attention will enhance
generalization without making the computationally unsustainable to be deployed in a real-world clinical setting
is not fully studied yet.
[25]
1.7 Novel contributions of this work
In order to fill the gaps, the work offers a transfer learning-based model of multi-classes of skin lesions
identification The summarized key contributions are as follows.
1. The paper uses an EfficientNet-B3 backbone as the main feature extractor of both dermoscopic and clinical
images of skin lesions. EfficientNet-B3 allows scaling the depth, width, and input resolution of the network to a
desirable trade-off between accuracy and computation costs, which is why it is appropriate with large datasets
in dermatology. Pre-training the network with weights trained on a huge collection of natural images effectively
transfers generic visual knowledge and eliminates the possibility of overfitting on small labelled but medical
data.
[26]
2. The work incorporates the Convolutional Block Attention Module (CBAM) in certain steps of EfficientNet-B3
to carry out combined channel and spatial attention. CBAM also improves intermediate feature maps through
modeling channel-wise importance by global pooling and gating, and by training on spatial attention masks
which emphasize important regions in the lesion. The inherent dual attention mechanism enables the network
to draw attention to diagnostically significant features, i.e. irregular pigment networks, unusual vascular
patterns and lesion boundary, and suppress background noise and artifacts.
[27]
3. The suggested architecture is trained on the DermNet-23 dataset (23 different classes of skin diseases), using
the Adam optimizer and the categorical cross-entropy loss. The pipeline of the training process consists of the
systematic preprocessing, the class-aware data augmentation, and the class-weighted optimization (where
applicable) to reduce the impact of the class imbalance. This model is compared on a set of extensive measures
which are the overall accuracy, macro and micro precision, recall, F1-score and per-class and macro-averaged
area under ROC curve which is a rigorous measure of the performance of the model on all classes.
[28]
4. The research also performs the comparative experiment with the baseline CNNs and a simple model of
EfficientNet-B3 without integrating CBAM. Such comparisons measure the value of the attention mechanism
and show that EfficientNet-B3 + CBAM setup becomes relatively stable in improving the macro-averaged F1-
score and AUC under the multi-class condition. The findings reveal the importance of the use of attention to
refine features in the distribution of complex dermatological images. Lastly, the paper focuses on the
reproducibility and clinical relevance of the work by describing the model structure, training plan and
evaluation protocol in a way that can be replicated and expanded by other researchers. The given framework
demonstrates how the efficient and attention-enhanced transfer learning can facilitate further development of
automated skin lesion classification and become the powerful background of the new computer-aided diagnosis
systems designed to assist dermatologists and primary care providers in early skin disease detection.
[29]
2. Related work
2.1 Handcrafted featurebased methods
Initial methods of skin lesion classification were mainly based on handcrafted visual features with classical
machine learning pipelines. In Codella et al.
[30]
, there was the use of multi-view framework that combined the
color and texture descriptors with sparse coding in melanoma detection, which focused on segmentation-based
feature fusion in order to enhance the sensitivity, but manual designing of features restricted their use in
heterogeneous images with imaging conditions. The sparse coding architecture suggested by Barata et al.
[31]
,
which uses local binary patterns (LBP) and color histograms with the support of support vector machines
(SVMs) on ISIC datasets, proved to be resilient to the variations in illumination, but appeared to have difficulties
when massive intra-class variability had to be considered. The system introduced by Ganster et al.
[32]
performs
an extraction of the asymmetry, irregularity at the border, color change and diameter using the segmented
lesions and diagnostic prediction using k-nearest neighbors, with results obtained being clinically viable when
applied to early dermoscopic datasets. Abbas et al.
[33]
developed hybrid methodologies, where the grey-level co-
occurrence matrix (GLCM) descriptors are coupled with wavelet-based color features and random forest
classifiers to achieve specificity in detecting basal cell carcinoma, but with preprocessing accuracy that is very
strict. Maglogiannis and Doukas
[34]
designed a mobile-specific pipeline based on principal component analysis
(PCA) with shape and chromatic features handcrafted followed by SVM classification, which allowed real-time
screening on consumer devices with lower accuracy on heterogeneous clinical imagery.
2.2 Classical CNNs and transfer learning
Deep convolutional neural networks have brought lesion analysis to end-to-end learning of features. In their
study, Esteva et al.
[35]
trained transfer learning with Inception-v3 on 129,450 clinical images to differentiate
keratinocyte carcinomas and seborrheic keratoses, with large-scale augmentation and ensemble optimization
to perform as well as a dermatologist would. Haenssle et al.
[36]
optimized ResNet-152 to melanoma detection on
the ISIC 2017 benchmark, including dermatologist-inspired preprocessing and test-time augmentation and
weighed the result at 86.5% AUC, a result superior to that of experts. Inception-v4 and ResNet-152 ensembles
were studied by Tschandl et al.
[37]
in a large-scale multicenter experiment and found the superiority of CNN in
binary diagnostic task with a generalized human advantage on a rare case. Asriani et al.
[38]
proposes a
technology-based solution by classifying skin cancer using a convolutional neural network (CNN) with a
ResNet50 architecture implemented into a mobile application via a REST API using Flask. Daneshjou et al.
[39]
performed a thorough study on transfer learning cases of VGG, ResNet and DenseNet backbones emphasizing
the necessity of domain adaptation between clinical and dermoscopic domains and uncovering the still
apparent shortcomings in multi-class generalization.
2.3 EfficientNet architectures for skin lesions
EfficientNet architectures introduced compound scaling to balance network depth, width, and spatial
resolution. Manole et al.
[40]
demonstrated the implementation of a custom model based on EfficientNetB3 has
demonstrated substantial potential for enhancing the diagnosis of skin lesions. This mode achieved a notably
high accuracy rate (95.4%/88.8%), underscoring the critical role of a comprehensive and diverse dataset.
Gessert et al.
[41]
fine-tuned EfficientNet-B4 for ISIC 2019 melanoma classification using pseudo-labeling and
test-time augmentation, achieving 0.915 AUC while reducing inference time relative to deep ResNet ensembles.
Chaturvedi et al.
[42]
applied EfficientNet-B3 to HAM10000 multi-class classification with lesion cropping and
class rebalancing, reporting 85.2% accuracy and demonstrating improved extraction of subtle dermoscopic
et al.
[43]
integrated EfficientNet-B0 with capsule networks for a 7-class
diagnostic model, achieving 95.6% accuracy on augmented DermNet subsets through hybrid attention fusion.
Huang et al.
[44]
modified EfficientNet-B5 for federated teledermatology during COVID-19, achieving a 93.8% F1-
score on diverse clinical images and validating feasibility for edge-device deployment.
2.4 Attention mechanisms in dermatological CNNs
Attention mechanisms have significantly enhanced CNN discriminability by prioritizing lesion-relevant regions.
Ocal
[45]
presented V-shaped network combining Spatial and channel squeeze-excitation (scSE) and edge
attention modules is proposed to enhance channelspatial focus and lesion boundary retention in skin lesion
segmentation. The model achieves superior performance, especially in IoU, on challenging ISIC datasets despite
hardware limitations. Mahbod et al.
[46]
integrated SE modules into multi-scale ResNets for ISIC lesion analysis,
reporting a 2.3% AUC improvement by emphasizing pigment-related features. Su et al.
[47]
introduced the
Convolutional Block Attention Module (CBAM). Shetty et al.
[48]
embedded in DenseNet-121 for HAM10000
multi-class classification, attaining a 4.1% macro-F1 improvement via artifact suppression and lesion-centered
focus. Qian et al.
[49]
proposed a grouping of multi-scale attention blocks (GMAB) which introduces different scale
attention branch to expand the DCNN model. Hanum et al.
[50]
combined channel and spatial attention within a
hybrid CNN-transformer architecture for 39-class lesion analysis, achieving 89.7% accuracy through cross-
attention fusion. Codella et al. surveyed attention-integrated architectures, including SECBAM hybrids,
documenting 37% sensitivity improvements for melanoma detection while identifying the need for expanded
multi-class evaluations across 20+ diagnostic categories.
2.5 Research Gaps
Despite advancements in attention-enhanced EfficientNet systems, several gaps remain. Prior studies such as
Gessert et al.
[41]
and Haenssle et al.
[36]
focus predominantly on binary melanoma detection, limiting applicability
to broader dermatological taxonomies such as DermNet-23's 23-class distribution. CBAM-based enhancements
(e.g., Shetty et al.
[48]
, Poma et al.
[50]
) improve macro-F1 performance but omit ablation studies comparing plain
EfficientNet-B3 baselines under severe class imbalance. Multi-scale attention frameworks (e.g., Qian et al.
[49]
)
improve recall yet lack macro-averaged per-class AUC reporting, especially for rare disorders. Although
handcrafted pipelines (e.g., Barata et al.
[32]
) remain valuable for interpretability, they do not match the
representational capacity of modern end-to-end architectures. Comprehensive evaluations unifying
EfficientNet-B3 with CBAM, supported by stratified metrics across the DermNet-23 dataset, channel- and
spatial-attention ablations, and edge-deployment feasibility analyses remain underexplored. These limitations
2.6 Summary
As summarized in Table 1, existing studies predominantly emphasize binary melanoma detection or limited
multi-class settings, often neglecting the challenges posed by large-scale dermatological taxonomies and severe
class imbalance. Although the recent attention-based methods exhibit better discriminative performance, they
often do not include systematic ablation studies, macro-averaged AUC analysis, or testing with clinically
heterogeneous datasets like the DermNet-23 one. It is in these gaps that there is a need to have a unified,
attention-directed, and parameter-efficient framework that can be able to classify skin lesion classified into
multiple classes in a robust manner; a fact that can be achieved through the given work.
3. Proposed methodology
The suggested approach will combine the state-of-the-art deep learning improvements to enhance the multi-
class skin lesion classification on mixed dermoscopic and clinical images. Based on EfficientNet-B3 as the main
feature extractor, the framework uses the elements of Convolutional Block Attention Module (CBAM) to increase
the feature discrimination of the channel and spatial, especially when dealing with minority and visually unclear
classes. The pipeline incorporates uniform preprocessing, lesion-focused augmentation, balanced training
plans, and systematic ablation to determine the role played by the attention mechanisms. This part describes
the model structure, data pre-treatment, training scheme and test procedure embraced to obtain strong and
generalizable classification results. Fig. 2 shows Flow diagram of the proposed EfficientNet-B3 + CBAM-based
skin lesion classification framework.
Table 1: Summary of existing studies, key contributions, and identified research gaps.
Sr. No
Study
Dataset / Task
Key Contributions
Limitations / Research Gaps
Ref.
1
Gessert et al.
(2020)
ISIC 2019,
Melanoma detection
Multi-resolution
EfficientNet ensemble with
metadata improves AUC
Limited to binary melanoma
classification; no evaluation on
large multi-class dermatology
datasets
[41]
2
Haenssle et al.
(2018)
ISIC 2017,
Melanoma vs benign
CNN performance
compared with
dermatologists
Focused on binary diagnosis;
lacks scalability to 23+ class
taxonomies
[36]
3
Esteva et al.
(2017)
Clinical images,
binary tasks
Achieved dermatologist-
level accuracy using
transfer learning
Does not address class
imbalance or fine-grained
multi-class differentiation
[35]
4
Tschandl et al.
(2020)
Multi-center
dermoscopy
HumanAI collaboration
improves performance
Primarily evaluates binary
tasks; limited per-class analysis
[37]
5
Shetty et al. (2020)
HAM10000, multi-
class
CNN-based dermoscopic
lesion classification
No attention ablation; limited
discussion on minority classes
[48]
6
Hanum et al.
(2025)
39-class dataset
Attention-guided deep
learning improves macro-
F1
Lacks baseline EfficientNet-B3
comparison and computational
cost analysis
[50]
7
Qian et al (2022)
HAM10000 dataset
The grouping of multi-scale
attention blocks (GMAB) to
extract multi-scale fine-
grained features
limitations in Sensitivity, Need
optimize the classification
accuracy of a small
number of classes.
[49]
8
Chaturvedi et al.
(2023)
HAM10000, multi-
class
EfficientNet-B3 with
ensemble improves
accuracy
No explicit attention
mechanisms; limited
interpretability analysis
[42]
9
Harahap et al.
(2024)
Dermoscopic images
EfficientNet architectures
outperform classical CNNs
Absence of attention modules
and ablation studies
[28]
10
Ul Amin et al.
(2024)
Video anomaly
datasets
EfficientNet + CBAM
improves feature
discrimination
Not designed for skin lesion
classification; domain
mismatch
[29]
11
Barata et al.
(2012)
Dermoscopy,
handcrafted features
Interpretable texture-color
features
Handcrafted pipelines lack
representational capacity of
modern DL
[32]
12
Ganster et al.
(2000)
Early dermoscopy
Rule-based automated
melanoma recognition
Not scalable to modern
datasets; outdated features
[33]
13
Maglogiannis &
Doukas (2008)
Mobile dermatology
Early computer-vision-
based screening
Reduced accuracy on
heterogeneous clinical images
[34]
3.1 Dataset and data preprocessing
The DermNet-23 dataset, comprising 15,557 RGB images across 23 dermatological disease classes, forms the
foundation for model development. Images exhibit variable resolutions (100×100 to 1024×768 pixels),
acquisition artifacts (hair, rulers, markers), and clinical heterogeneity reflecting real-world dermoscopy and
photography conditions. Preprocessing ensures input consistency for EfficientNet-B3: (1) resizing to 300×300
.485, 0.456, 0.406],
-hat filtering and inpainting; (4)
contrast-limited adaptive histogram equalization (CLAHE, clip limit=2.0) for lesion enhancement; and (5)
Dataset Preprocessing EfficientNet-B3
Output Classifier
CBAM
optional lesion-centric cropping using Otsu thresholding where segmentation masks are available. These steps
mitigate domain shift and background noise while preserving diagnostically relevant textures and pigment
patterns.
3.2 Data Splitting
Class distributions are kept in stratified splitting: 70% training (10,890 images), 15% validation (2,334 images)
and 15% test (2,333 images), so that each class has 30 or more samples in the validation/test sets, which is
sufficient to evaluate the macro-averaged performance. Hyperparameter optimization is achieved by using five-
fold stratified cross-validation on the training/validation split (80/20) to avoid overfitting and give unbiased
estimates of generalization to the held-out test set. Class imbalance is directly monitored through verification
of per-fold minority class sampling.
3.3 Model training
The model training phase incorporates the transfer learning, attention, and supervised optimization algorithms
to create a powerful classifier in the identification of the lesion on the skin in multi-classes. The suggested
method uses the EfficientNet-B3 as the main feature extractor and complements it with the Convolutional Block
Attention Module (CBAM) to enhance the discriminative power of the learnt representations.
3.4 Transfer learning with EfficientNet-B3
EfficientNet-B3 is chosen because of its scaling strategy of compounds, which optimizes the depth, width and
input resolution of the network at the same time. It is also decentralized in that it is started with ImageNet-
pretrained weights allowing it to exploit generalized low-level and mid-level visual features, which greatly
accelerates convergence and extends model behaviour to outliers in medical imaging tasks. In order to stabilize
and adapt flexibly, the first convolutional layers remain frozen throughout the early steps of the training process,
and more layers are gradually unfrozen to enable fine-tuning of lesion-specific patterns. This is a training
strategy with stages that helps to reduce the risk of catastrophic forgetting as well as stabilize gradient flow.
Fig. 3: The Convolutional Block Attention Module (CBAM).
Fig. 3 shows The Convolutional Block Attention Module (CBAM). CBAM is incorporated into selected
EfficientNet-B3 blocks to enhance feature refinement. CBAM operates through two sequential attention
mechanisms:
• Channel Attention: Trains a weight of each channel in the feature. To compute statistics through global
average pooling and max pooling, squeeze and excitation operations are computed and shared multi-layer
perceptrons (MLPs) are used to produce channel-wise attention maps. The attentions maps put a focus on
clinically informative textures and pigmentations patterns.
• Spatial Attention: Plays attention to the location of the most discriminative aspects of the lesion. The
computation of the spatial attention maps consists of the convolution of the pooled channel descriptors,
which emphasize images areas that can be characterized as asymmetry, irregularity of the border or
abnormal pigmentation.
CBAM together with the network helps in placing emphasis on the medically relevant structures and reducing
irrelevant background noise.
3.5 Classification head
After attention-enhanced feature extraction, the output is fed into a classification head consisting of:
• Global Average Pooling (GAP)
• Dropout (rate = 0.3) to reduce overfitting
• A dense layer with ReLU activation
• A final SoftMax layer producing probability scores for the 23 lesion classes
The design is such that it gives a small but expressive output representation, which can be used in multi-class
classification.
3.5.1 Optimization Algorithm
The Adam optimizer is applied to learn and train stably and efficiently with a combination of adaptive learning
rates and momentum. Medical imaging work is especially well done with Adam, as it supports the
heterogeneous feature distributions. The optimizer optimizes the model parameters based on:
(1)
The bias-corrected estimates of the first and second moment of the gradient, m t and v t, respectively, are known
as the bias-corrected first and second moment, respectively, and the learning rate is 1/alpha. It uses a cosine
decaying scheduler that decreases the learning rate with the number of epochs to allow the initial stages of
training to learn coarsely and then the subsequent stages of training to fine-tune the parameter.
3.6 Loss function and class imbalance handling
The network is optimized using the Categorical Cross-Entropy loss function, defined as:
(2)
yi is the ground-truth class label and y i is the probability of classiii.
Considering the uneven classification of DermNet-23 data, the weights of classes are added to the loss to
penalize the mistake of the minority classes with an extra burden. This change makes underrepresented lesion
types more sensitive and makes the classification more robust in general.
3.7 Training strategy
The training pipeline is a multi-stage process that is structured to have a stable convergence and successful
feature learning:
1. Freeze the initial EfficientNet-B3 layers in order to retain pretrained low-level feature representations.
2. Training the classifier head only under a warming up period to stabilize the gradient flow.
3. Layers of the backbone progressively unfreezing to allow finetuning on lesion patterns.
4. Attention-enhanced representation learning
in the full network with the integrated CBAM modules.
5. Checking the validation loss to implement early stopping where needed.
6. Retaining the best performing model checkpoint on the validation performance.
This staged optimization approach improves the stability of training, avoids overfitting, and facilitates good
quality of discriminative feature representation of all lesion types.
3.8 Evaluation metrics
The performance measurement uses overall metric which cures multi-class imbalance: the aggregate accuracy,
macro/micro-average precision (P), recall (R), F1-score and area under the ROC curve (AUC). In macro
averaging, classes are given equal weight, giving more weight to rare conditions; in micro averaging, there is a
weight on overall accuracy in prediction. Per-class measures the failure modes (e.g. melanoma vs. nevi
confusion) and confusion matrices are used to visualize the patterns of errors. The kappa is an agreement
measure created by Cohen that indicates whether there is agreement on a medical condition or not and top-3
accuracy is an evaluation of clinical utility that involves a scenario in which dermatologists will look at
differentials. The McNemar test (p<0.05) is used to test statistical significance across cross-validation folds.
In order to measure the performance in a comprehensive way, several quantitative measures were calculated:
1. Accuracy
(3)
Represents the overall proportion of correctly classified samples.
2. Precision
(4)
Measures the reliability of positive predictions, i.e., how many predicted disease cases are correct.
3. Recall (Sensitivity)
(5)
4. F1-Score
(6)
Provides a harmonic mean of precision and recall, balancing both in a single metric.
5. AUC and ROC analysis
Each class is calculated in a one-vs-rest fashion Area Under the Receiver Operating Characteristic Curve (AUC),
and averaged to give a macro-AUC. AUC is the degree of separability of correct and incorrect prediction at
different thresholds, less insensitive to the imbalance of the two classes than accuracy.
(7)
A higher AUC indicates improved discrimination between lesion categories.
6. Confusion matrix
A multi-class confusion matrix is generated to visualize classification behaviour for each lesion type. It
highlights:
• Misclassification patterns
• Confusion between clinically similar categories
• Improvements resulting from CBAM attention mechanisms
This provides actionable insights for refining the model.
7. Ablation studies
To assess the contribution of attention modules, evaluation metrics are compared across:
1. Baseline EfficientNet-B3
2. EfficientNet-B3 + Channel Attention
3. EfficientNet-B3 + Spatial Attention
4. EfficientNet-B3 + CBAM (full attention)
Ablation analysis measures the performance improvement which can be attributed to each design element.
Accuracy, macro-F1, macro-AUC and confusion matrix visualization work well together as a set of evaluation.
This makes sure that the proposed model is accurate, as well as reliable when applied to all 23 lesion classes,
including rare and visually ambiguous classes.
3.9 Testing
The held-out test set is then evaluated with the ensemble-averaged model (5 cross-validation folds) using the
same preprocessing (no augmentation). Inference uses test-time augmentation (10 crops per image, average
horizontal folds) and soft-voting folds to make robust predictions. Latency is captured on RTX 3050 (minimum
50ms/image to be used clinically). Gradient-weighted class activation maps (Grad-CAM) visualization maps the
focus of attention, confirming its localization in the areas of the lesions, rather than in the background. Subsets
of ISIC-2018 external validation mimic domain shift, which measures generalization to unseen dermoscopic
data.
4. Result and discussion
This section presents a comprehensive evaluation of the proposed EfficientNet-B3 + CBAM architecture on the
DermNet-23 dataset. The performance of the model is analyzed quantitatively using standard classification
metrics and qualitatively through confusion matrix visualizations and Grad-CAMbased interpretability
assessments. Comparative experiments with baseline architectures demonstrate the effectiveness of the
proposed approach in addressing class imbalance, visual ambiguity, and multi-class complexity.
4.1 Result of proposed EfficientNet-B3 + CBAM
As indicated in the training and validation curves, the proposed EfficientNet-B3 + CBAM model is effective in
learning and good at generalizing. The loss curve of both the loss curve monotonically decreases and validation
loss remains low relative to training loss, which shows that there is good regularization and good optimization.
The accuracy curves also indicate that validation accuracy is greater than training accuracy in all the epochs,
which indicate that the model is more effective on clean validation data compared to the augmented training
set. This behavior confirms that the network is neither overfitting nor underfitting and is successfully capturing
discriminative features from the input images.
Fig. 4(a) and Fig. 4(b) present the training and validation accuracy and loss curves of the proposed EfficientNet-
B3 + CBAM model on the DermNet-23 dataset. The accuracy of the training shifts slowly to a steady level of 63-
65, and the accuracy of the validation is rapidly growing to a steady level of about 76-78 and is a sign of
successful learning and good generalization. In line with this, the initial training loss of above 3.0 drops to an
almost 1.25 but the validation loss drops more and approaches 0.95-1.0. The gradual and gradual decline in the
loss and the near perfect correspondence between the training and the validation curve verify that convergence
was stable and that the optimization was successful and that there were minimal overfitting even in the case of
the extreme class imbalance and multi-class characteristics of the DermNet-23 dataset. These remarks show
that the compound scaling approach of EfficientNet-B3 with attention-focused feature optimization allows
learning discriminative dermatological patterns efficiently.
Fig. 4(c) and Fig. 4(d), reflect the performance of the proposed model in classification using ROC and confusion
matrix. The AUC of the one- vs-rest ROC curve is macro averaged at 0.94, which indicates that the ROC was
strong in its discriminative ability across all 23 disease categories. A high level of diagonal dominance with a
minor misclassification between the conditions of distinct clinical interest can be seen in the confusion matrix.
This tendency underlines the performance of the CBAM module in focusing attention on salient spatial areas
and informative channel characteristics, which leads to the balanced performance with the overall classification
accuracy of 87.1, macro-average precision of 86.2, macro-average recall of 85.0 and macro-F1 score of 85.6 as
indicated in Table 2.
Table 2: Performance metrics derived from aggregated confusion matrix statistics.
Metric
Formula
Value
Accuracy
(TP + TN) / Total Samples
87.1%
Precision (Macro)
TP / (TP + FP)
86.2%
Recall (Macro)
TP / (TP + FN)
85.0%
F1-Score (Macro)
2PR / (P + R)
85.6%
AUC (Macro)
One-vs-rest ROC area
0.94
To give a quantitative analysis of the visual confusion matrix, aggregated confusion matrix statistics (Table 3)
and calculated evaluation metrics (Table 2) are summarized.
Table 3: Aggregated confusion matrix statistics for the proposed EfficientNet-B3 + CBAM Model.
Metric
Value
True Positives (TP)
1983
False Positives (FP)
317
False Negatives (FN)
350
True Negatives (TN)
Not uniquely defined (multi-class setting)
Total Test Samples
2333
Due to the multi-class (23-class) nature of the problem, TP, FP, and FN are reported in an aggregated manner,
while TN is not uniquely defined and therefore is not reported as a scalar value. The combined confusion matrix
statistics and macro-averaged measures substantiate the statement that the proposed EfficientNet-B3 + CBAM
model demonstrates balanced and strong classification results in all lesion categories in the case of class-
imbalanced conditions.
4.2 Qualitative and quantitative analysis
The performance of the proposed model in terms of accuracy, macro-averaged precision, recall, F1-score, and
area under the receiver operating characteristic curve (AUC) are used as standard multi-class classification
metrics to measure the quantitative aspect of performance of the proposed EfficientNet-B3 + CBAM model. The
model in question has an overall accuracy of 87.1 which is significantly higher than those of Classical CNN,
ResNet50, MobileNetV3 and plain EfficientNet-B3 architectures. The macro-averaged precision of 86.2% and
recall of 85.0% reveal equal predictive ability with majority classes and minority classes whereas the macro-F1
of 85.6% reveals strong harmonic performance in case of class imbalance. Moreover, the proposed strategy
achieves a macro-AUC of 0.94 indicating high levels of class separability when analysing one-vs-rest ROC and
high levels of discriminative ability in all the dermatologic categories evaluated.
Fig. 4: The result of proposed EfficientNet-B3 + CBAM (a) Accuracy, (b) Loss (c) ROC Curve (d) Confusion Matrix.
Table 4 presents the summaries of the comparative quantitative performance of the proposed model with four
baseline architectures, including Classical CNN, ResNet50, MobileNetV3, and EfficientNet-B3. The findings
indicate the steady positive changes across all assessment measures and confirm the efficiency of attention-
guided refinement of features provided by the CBAM module in the combination with the EfficientNet-B3
backbone.
Qualitative analysis is done by visual analysis of the confusion matrix and ROC curves in Fig. 4. The confusion
matrix of the proposed EfficientNet-B3 + CBAM model has a high diagonal dominance, which means that
numerous samples of lesions are accurately identified. Inconsistent classifications occur in majority of the
clinically similar dermatological groups including visually similar inflammatory and pigmented lesions, and are
attributed to realistic diagnostic uncertainty, and not haphazard prediction mistakes. Such organized error
distribution is in line with the achieved macro-F1 score and AUC improvement.
Table 4: Performance comparison of baseline models and the proposed EfficientNet-B3 + CBAM Architecture.
Model
Accuracy
(%)
Precision
(%)
Recall
(%)
F1-Score
(%)
AUC
Classical CNN
72.5
69.8
69.8
69.8
0.84
ResNet50
78.2
76.5
76.5
76.5
0.88
MobileNetV3
80.1
78.3
78.3
78.3
0.89
EfficientNet-B3
82.4
80.7
80.7
80.7
0.91
Proposed EfficientNet-B3 + CBAM
87.1
86.2
85.0
85.6
0.94
ROC analysis of one- vs-rest is also used to test the discriminative ability of the considered models. The proposed
EfficientNet-B3 + CBAM model records the best macro-averaged AUC of 0.94 compared to EfficientNet-B3 (AUC
= 0.91) and the Classical CNN (AUC = 0.84). The ROC curves show that the true positive rates are high and the
false positive rates do not inflate when the decision threshold varies, which makes the existence of strong multi-
class separability. Overall, the quantitative output confirms the existence of evident performance improvement
of the proposed architecture, whereas the qualitative analysis gives insight into the classification behavior and
error characteristics of the proposed architecture. All these results show that incorporating EfficientNet-B3
along with CBAM creates the powerful and clinically useful multi-class skin lesion classification framework,
which can successfully address class imbalance and visual similarity due to different dermatology diseases.
5. Conclusion and future scope
This was a paper where a deep learning model to classify skin lesions (several classes) was presented,
combining EfficientNet-B3 with the Convolutional Block Attention Modules (CBAM). The suggested
methodology used systematic preprocessing, balanced data augmentation and attention-based feature
refinement to overcome major issues like imbalance of classes, visual diversity and subtle inter-class similarities
that characterize large-scale dermatological dataset like DermNet-23. Experimental analysis proved that the
suggested EfficientNet-B3 + CBAM model had the overall classification accuracy of 87.1, macro-averaged
precision of 86.2, macro-averaged recall of 85.0, and macro-F1 score of 85.6, which proved the balanced and
reliable performance in both the majority and minority lesion categories. Moreover, the macro-averaged AUC of
the model is 0.94, which indicates the high-class separability when using a one-vs-rest ROC analysis.
Comparative experiments proved the attention-augmented architecture was always better than Classical CNN,
ResNet50, MobileNetV3, and plain EfficientNet-B3 baselines, and proved that channel spatial attention is
effective to improve the discriminative features learning. The studies on ablation further supported the personal
and joint efforts of the components of CBAM and showed that the components were more robust and could be
generalized without important computational demands. Even with these encouraging outcomes, various areas
of research can also be used to increase the clinical relevance and performance of the proposed system. To be
accurate, generalization to other skin tones, other imaging devices, and acquisition conditions, such as in the
real world, can be enhanced by extending training and evaluation to larger and more diverse and annotated
datasets. Second, the use of attention mechanisms based on transformers, self-supervised pretraining methods,
as well as multimodal fusion (using a dermoscopic image with clinical metadata) can be further adopted to
enhance diagnostic accuracy and robustness. Third, privacy preserving model adaptation in distributed medical
institutions can be facilitated by the integration of federated learning frameworks. Also, the inclusion of
explainability methods including Grad-CAM++ or attention-based saliency maps can increase clinical decision-
support transparency and trust. Lastly, the real-time deployment of the model on lightweight edge devices, as
well as its optimization and the ability to engage in constant improvements through active learning, are also
promising directions of building scalable, reliable, and accessible dermatological AI systems.
References
[1]
M. Wang, X. Gao, L. Zhang, Recent global patterns in skin cancer incidence, mortality, and prevalence,
Chinese Medical Journal, 2025, 138, 185-192, doi: 10.1097/CM9.0000000000003416.
[2]
World Health Organization, Skin cancers, 2023, WHO Fact Sheet. https://www.who.int
[3]
M. Arnold, D. Singh, M. Laversanne, J. Vignat, S. Vaccarella, F. Meheus, A. E. Cust, E. de Vries, D. C.
Whiteman, F. Bray, Global burden of Cutaneous Melanoma in 2020 and Projections to 2040. JAMA
Dermatology, 2022, 158, 495503. doi: 10.1001/jamadermatol.2022.0160.
[4]
A. H. Roky, M. M. Islam, A. M. Fuad Ahasan, Md. S. Mostaq, Md. Z. Mahmud, M. Nurul Amin, Md. Ashiq
Mahmud, Overview of skin cancer types and prevalence rates across continents, Cancer Pathogenesis
and Therapy, 2025, 3, 89-100, doi: 10.1016/j.cpt.2024.08.002.
[5]
X. Wu, M. A. Marchetti, A. A. Marghoob, Dermoscopy: Not just for Dermatologists. Melanoma
Management, 2015, 2, 6373, doi: 10.2217/mmt.14.32.
[6]
M. Zortea, T. R. Schopf, K. Thon, M. Geilhufe, K. Hindberg, H.Kirchesch, K. Møllersen, J. Schulz, S. Olav
Skrøvseth, F. Godtliebsen, Performance of a dermoscopy-based computer vision system for the
diagnosis of pigmented skin lesions compared with visual evaluation by experienced dermatologists,
Artificial Intelligence in Medicine, 2014, 60, 13-26, doi: 10.1016/j.artmed.2013.11.006.
[7]
K. Liopyris, S. Gregoriou, J. Dias, A. J. Stratigos, Artificial intelligence in dermatology: Challenges and
perspectives, Dermatology and Therapy, 2022, 12, 26372651, doi: 10.1007/s13555-022-00833-8.
[8]
L. Wang, L. Zhang, X. Shu, Z. Yi, Intra-class consistency and inter-class discrimination feature learning
for automatic skin lesion classification, Medical Image Analysis, 2023, 85, 102746, doi:
10.1016/j.media.2023.102746.
[9]
A. Akram, J. Rashid, M. A. Jaffar, M. Faheem, R. U. Amin, Segmentation and classification of skin lesions
using a hybrid deep learning method in the Internet of Medical Things, Skin Research and Technology,
2023, 29, e13524. doi: 10.1111/srt.13524.
[10]
A. Murugan, S. Anu H Nair, A. Angelin Peace Preethi, K. P. Sanal Kumar, Diagnosis of skin cancer using
machine learning techniques, Microprocessors and Microsystems, 2021, 81, 103727, doi:
10.1016/j.micpro.2020.103727.
[11]
N. V. Kumar, P. V. Kumar, K. Pramodh, Y. Karuna, Classification of skin diseases using Image processing
and SVM, 2019 International Conference on Vision Towards Emerging Trends in Communication and
Networking (ViTECoN), Vellore, India, 2019, 1-5, doi: 10.1109/ViTECoN.2019.8899449.
[12]
M. Q. Hatem, Skin lesion classification system using a K-nearest neighbor algorithm, Visual Computing
for Industry, Biomedicine, and Art, 2022, 5, doi: 10.1186/s42492-022-00103-6.
[13]
S. Mustafa, A. Jaffar, M. Rashid, S. Akram, S. M. Bhatti, Deep learning-based skin lesion analysis using
hybrid ResUNet++ and modified AlexNet-Random Forest for enhanced segmentation and
classification, doi: 10.1371/journal.pone.0315120
[14]
P. Yang, Z. Chen, X. Sun, X. Deng, Better with less: efficient and accurate skin lesion segmentation
enabled by diffusion model augmentation, Electronics, 2025, 14, 3359, doi:
10.3390/electronics14173359.
[15]
A. Abobakir, A. Abdulazeez, A review on utilizing machine learning classification algorithms for skin
cancer. Journal of Applied Science and Technology Trends, 2022, 5, 6071, doi: 10.38094/jastt52191.
[16]
A. Toprak, I. Aruk, A hybrid convolutional neural network model for the classification of multi-class
skin cancer., International Journal of Imaging Systems and Technology, 2024, 34, e23180, doi:
10.1002/ima.23180.
[17]
A. S. Al-Waisy, S. Al-Fahdawi, M. I. Khalaf, M. A. Mohammed, B. Al-Attar, M .N. Al-Andoli, A deep learning
framework for automated early diagnosis and classification of skin cancer lesions in dermoscopy
images, Scientific Reports, 2025, 15, 31234, doi: 10.1038/s41598-025-15655-9.
[18]
M. A. H. Lubbad, I. L. Kurtulus, D. Karaboga, K. Kilic, A. Basturk, B. Akay, O. U. Nalbantoglu, O. M. Durmaz
Yilmaz, M. Ayata, S. Yilmaz, I. Pacal, A comparative analysis of deep learning-based approaches for
classifying dental implants decision support system, Journal of Imaging Informatics in Medicine, 2024,
37, 25592580, doi: 10.1007/s10278-024-01086-x.
[19]
F. Brutti, F. La Rosa, L. Lazzeri, C. Benvenuti, G. Bagnoni, D. Massi, M. Laurino, Artificial intelligence
algorithms for Benign vs. Malignant Dermoscopic skin lesion image classification, Bioengineering,
2023, 10, 1322. doi: 10.3390/bioengineering10111322.
[20]
H. Hussein, A. Magdy, R. F. Abdel-Kader, K. Abd El Salam, Binary classification of skin cancer images
using pre-trained networks with I-GWO. Inteligencia Artificial, 2024, 27, 102116, doi:
10.4114/intartif.vol27iss74pp102-116.
[21]
A. Kalaivani, S. Karpagavalli, Detection and classification of skin diseases with ensembles of deep
learning networks in medical imaging, International Journal of Health Sciences, 2022, 1362413637,
doi: 10.53730/ijhs.v6ns1.8402.
[22]
M. Tan, Q. V. Le, EfficientNet: Rethinking model scaling for convolutional neural networks. In
Proceedings of the 36th International Conference on Machine Learning, 2019, 61056114.
[23]
A. A. Abd El-Aziz, M. A. Mahmood, S. A. El-Ghany, EfficientNet-B3-based automated deep learning
framework for multiclass endoscopic bladder tissue classification, Diagnostics, 2025, 15, 2515, doi:
10.3390/diagnostics15192515.
[24]
K. Kanchana, S. Kavitha, K. J. Anoop, B. Chinthamani, Enhancing skin cancer classification using
EfficientNet B0B7 through transfer learning, Asian Pacific Journal of Cancer Prevention, 2024, 25,
17951802, doi: 10.31557/APJCP.2024.25.5.1795.
[25]
J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, IEEE, 2018, 71327141.
[26]
S. Woo, J. Park, J.-Y. Lee, I. S. Kweon, CBAM: Convolutional block attention module, In Proceedings of the
European Conference on Computer Vision, Sprigger, 2028, 319, doi:
https://arxiv.org/abs/1807.06521.
[27]
M. Harahap, J. Leonardi, S. C. Kwok, D. M. Ong, A. M. Husein, D. Ginting, B. A. Silitonga, V. Wizley, Skin
cancer classification using EfficientNet architecture, Bulletin of Electrical Engineering and Informatics,
2024, 13, 27162728, doi: 10.11591/eei.v13i4.7159.
[28]
S. Ul Amin, Y. Jung, B. Kim, M. S. Abbas, S. Seo, Enhanced anomaly detection using EfficientNet and
CBAM, IEEE Access, 2024, 12, 162697162712, doi: 10.1109/ACCESS.2024.3488797.
[29]
S. Remya, T. Anjali, V. Sugumaran, A novel transfer learning framework for multimodal skin lesion
analysis. IEEE Access, 2024, 12, 5073850754, doi: 10.1109/ACCESS.2024.3385340.
[30]
N. Codella, V. Rotemberg, P. Tschandl, M. Emre Celebi, S. Dusza, D. Gutman, B. Helba, A. Kalloo, K.
Liopyris, M. Marchetti, H. Kittler, A. Halpern, Skin lesion analysis toward melanoma detection: A
challenge at the International Symposium on Biomedical Imaging (ISBI), hosted by the International
Skin Imaging Collaboration (ISIC). In Proceedings of the IEEE International Symposium on Biomedical
Imaging IEEE, 2018,168172.
[31]
J. Barata, M. Ruela, M. Francisco, T. Mendonça, J. Marques, Two systems for the detection of melanomas
in dermoscopy images using texture and color features. In Proceedings of the IEEE International
Symposium on Biomedical Imaging, IEEE, 2012, 4952.
[32]
H. Ganster, P. Pinz, R. Röhrer, E. Wildling, M. Binder, H. Kittler, Automated melanoma recognition, IEEE
Transactions on Medical Imaging, 2001, 20, 233-239, doi: 10.1109/42.918473.
[33]
Q. Abbas, I. F. Garcia, M. E.Celebi, W. Ahmad, A feature-preserving hair removal algorithm for
dermoscopy images, Skin Research and Technology, 2013, 19, e103e120, doi: 10.1111/srt.12028.
[34]
I. Maglogiannis, C. N. Doukas, Overview of advanced computer vision systems for dermatological
applications, International Journal of Artificial Intelligence Tools, 2008, 17, 921936, doi:
10.1142/S0218213008004368.
[35]
A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, S, Thrun, Dermatologist-level
classification of skin cancer with deep neural networks, Nature, 2017, 542, 115118, doi:
10.1038/nature21056.
[36]
H. A. Haenssle, C. Fink, R. Schneiderbauer, F. Toberer, T. Buhl, A. Blum, A. Kalloo, A. Ben Hadj Hassen,
L. Thomas, A. Enk, L. Uhlmann, C. Alt, M. Arenbergerova, R. Bakos, A. Baltzer, I. Bertlich, A. Blum, T.
Bokor-Billmann J. Bowling, N. Braghiroli, R. Braun, K. Buder-Bakhaya, T. Buhl, H. Cabo, L. Cabrijan, N.
Cevic, A. Classen, D. Deltgen, C. Fink, I. Georgieva, L. Hakim-Meibodi, S. Hanner, F. Hartmann, J.
Karls, H. Koga, J. Kreusch, A. Lallas, P. Majenka, A.Marghoob. C.
Massone, L. Mekokishvili, D. Mestel. V. Meyer, A. Neuberger, K. Nielsen, M. Oliviero, R. Pampena, J.
Paoli, E. Pawlik, B. Rao. A. Rendon, T. Russo. A. Sadek, K. Samhaber, R. Schneiderbauer, A. Schweizer, F.
Toberer. L. Trennheuser, L. Vlahova. A. Wald, J. Winkler, P. Wölbing I. Zalaudek, Man against machine:
Diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma
recognition in comparison to 58 dermatologists, Annals of Oncology, 2018, 29, 18361842, doi:
10.1093/annonc/mdy166.
[37]
P. Tschandl, C. Rinner, Z. Apalla, G. Argenziano, N. Codella, A. Halpern, M. Janda, A. Lallas, C. Longo, J.
Malvehy, J. Paoli, S. Puig, C. Rosendahl, H. Peter Soyer, I. Zalaudek, H. Kittler, Humancomputer
collaboration for skin cancer recognition, Nature Medicine, 2020, 27, 17, doi: 10.1038/s41591-020-
0942-0.
[38]
A. Asriani, N. T. Lapatta, D. W. Nugraha, A. Amriana, W. Wirdayanti, Implementation of ResNet-50-
based convolutional neural network for mobile skin cancer classification, Journal of Applied
Informatics and Computing, 2025, 9, 19691577, doi: 10.30871/jaic.v9i4.9696.
[39]
R. Daneshjou, K. Vodrahalli, R. A. Novoa, M. Jenkins, W. Liang, V. Rotemberg, J. Ko, S. M Swetter, E. E.
Bailey, O. Gevaert 2, P. Mukherjee, M. Phung, K. Yekrang, B. Fong, R. Sahasrabudhe, J. A. C Allerup, U.
Okata-Karigane, J. Zou, A. S. Chiou, Disparities in dermatology AI performance on a diverse, curated
clinical image set, NPJ Digital Medicine, 2021, 4, 156, doi: 10.1038/s41746-021-00511-2.
[40]
I. Manole, A.-I. Butacu, R. N. Bejan, G. -S. Tiplica, Enhancing dermatological diagnostics with
EfficientNet: A deep learning approach, Bioengineering, 2024, 11, 810, doi:
10.3390/bioengineering11080810.
[41]
N. Gessert, M. Nielsen, M. Shaikh, R. Werner, A. Schlaefer, Skin lesion classification using ensembles of
multi-resolution EfficientNets with meta data, MethodsX, 2020, 7, 100864, doi:
10.1016/j.mex.2020.100864.
[42]
S. S. Chaturvedi, K. K. Nagwanshi, S. Singh, Skin lesion analyser using modified AlexNet and
EfficientNet-B3. Biomedical Signal Processing and Control, 2023, 79, 104201, doi:
10.1016/j.bspc.2022.104201.
[43]
-based medical image classification. Medical Hypotheses, 2020,
135, 109833, doi: 10.1016/j.mehy.2019.109833.
[44]
S.-C. Huang, M.-Y. Pare, Y.-H. Chung, Y.-L. Tang, S.-A. Parsons, Integrated structure tensor and deep
neural network for melanoma segmentation, IEEE Access, 2020, 8, 2467924691, doi:
10.1109/ACCESS.2020.2970341.
[45]
H. Ocal, scSEETV-Net: Spatial and channel squeeze-excitation and edge attention guidance V-shaped
network for skin lesion segmentation. Advanced Intelligent Systems, 2024, 6, 2400438, doi:
10.1002/aisy.202400438.
[46]
A. Mahbod, G. Schaefer, C. Wang, R. Ecker, I. Ellinge, Skin lesion classification using hybrid deep neural
networks, ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Brighton, UK, 2019, 1229-1233, doi: 10.1109/ICASSP.2019.8683352.
[47]
Q. Su, H. N. A. Hamed, D. Zhou, Relation Explore Convolutional block attention module for skin lesion
classification, International Journal of Imaging Systems and Technology, 2024, 35, e70002, doi:
10.1002/ima.70002.
[48]
B. Shetty, R. Fernandes, A. P. Rodrigues, R. Chengoden, S. Bhattacharya, K. Lakshmanna, Skin lesion
classification of dermoscopic images using machine learning and convolutional neural network,
Scientific Reports, 2022, 12, 18134, doi: 10.1038/s41598-022-22644-9.
[49]
S. Qian, K. Ren, W. Zhang, H. Ning, Skin lesion classification using CNNs with grouping of multi-scale
attention and class-specific loss weighting, Computer Methods and Programs in Biomedicine, 2022, 226,
107166, doi: 10.1016/j.cmpb.2022.107166.
[50]
S. A. Hanum, A. Dey, M. S. Kabir, An attention-guided deep learning approach for classifying 39 skin
lesion types, Image and Video Processing, arxiv Preprint, doi: 10.48550/arXiv.2501.05991.
Publisher Note: The views, statements, and data in all publications solely belong to the authors and
contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products
mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional
affiliations.
Open Access
This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which
permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,
as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative
Commons License and changes need to be indicated if there are any. The images or other third-party material
in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line
to the material. If material is not included in the article's Creative Commons License and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-
nc/4.0/
© The Author(s) 2025
