Lung cancer remains among the leading causes of cancer-related mortality worldwide, with non-small cell lung cancer

(NSCLC) accounting for approximately 85% of all cases.

The early detection and accurate classification of NSCLC

are crucial for improving patient prognosis and tailoring effective treatment strategies. However, detecting and

classifying small nodules, particularly those less than 3 mm in size,

pose significant challenges because of the

further compounded when predicting the risk of recurrence, a critical factor in the long-term management of NSCLC

patients.

To address these challenges, the proposed research introduces a novel, interpretable framework that leverages

advanced deep learning techniques to enhance the detection, classification, and recurrence risk prediction of small

NSCLC nodules. Central to this framework is the integration of a CNN-Transformer hybrid model, which

synergistically combines the spatial feature extraction strengths of convolutional neural networks (CNNs) with the

long-range dependency modeling capabilities of transformers. This hybrid approach enables the model to capture

intricate features of small nodules with high precision, achieving a detection accuracy of 95% and a classification

accuracy of 93%. In addition to the hybrid model, the framework incorporates a generative adversarial network-based

super-resolution (SRGAN) technique to overcome the resolution limitations of standard CT scans.

temporal sequences of medical images, capturing the progression of the disease over time and providing a recurrence

risk prediction accuracy of 88%.

To ensure the interpretability of the model’s predictions, layer-wise relevance propagation (LRP) is applied, generating

heatmaps that highlight the critical regions influencing the predictions.

This interpretability aspect is vital for clinical

adoption, as it provides transparency and enhances the trust of medical professionals in the model's outputs. In

summary, this research presents a comprehensive and interpretable deep learning-based framework for early detection,

classification, and recurrence risk prediction of small NSCLC nodules. By addressing the limitations of existing

methods and integrating state-of-the-art techniques such as CNN-Transformer hybrids, the SRGAN, self-attention, and

intervention can significantly improve patient outcomes. However, current imaging techniques and computational

models exhibit limitations in terms of their resolution and feature extraction capabilities, which restrict their

effectiveness in identifying and classifying these minute nodules.

Furthermore, predicting the risk of recurrence in

NSCLC patients remains a formidable challenge, as existing models often fail to capture the complex temporal

dynamics and heterogeneity of tumor progression. These challenges necessitate the development of a novel, integrated

approach that not only enhances the detection and classification accuracy of small nodules but also provides reliable

and interpretable predictions of recurrence risk.

In this context, the proposed research makes several significant contributions to the field of medical imaging and

cancer prognosis. Some of the contributions are as follows:

NSCLC nodules. Additionally, the framework incorporates a generative adversarial network-based super-resolution

(SRGAN) technique to address the resolution limitations of traditional CT scans.

training, significantly improving the visualization and detection sensitivity for small nodules. The use of self-

attention mechanisms further enhances the model's ability to detect subtle features that are indicative of small

nodules by dynamically focusing on the most relevant regions of the image samples. This approach increases the

detection accuracy by allowing the model to prioritize critical areas during the feature extraction process.

medical images over time, capturing the temporal patterns associated with disease progression. To ensure the

forefront of medical research for years. Advances in imaging technologies, computational methods, and data analytics

have driven significant progress in this field. This review synthesizes findings from recent studies, highlighting the

diverse methodologies employed and their corresponding results, with an emphasis on both the achievements and

limitations of each approach. Through this synthesis, we aim to elucidate the current state of lung cancer research and

identify pathways for future advancements. As shown in Table 1, the studies under review encompass a variety of

techniques ranging from deep learning models and radiomics to multiple-modal data integration and genetic analysis.

For instance, the work by Ghita et al.

focused on parameterizing respiratory impedance in lung cancer patients using

forced oscillation lung function tests. This approach improved the accuracy of lung function tests, yet its applicability

was limited to the specific technique employed. In contrast, Wu et al.

utilized multiple-view adaptive weighted graph

to improved accuracy in identification and characterization. However, this method is heavily dependent on the quality

of positron emission tomography (PET) or CT images and samples. Similarly, Tortora et al.

learning approaches for adaptive radiotherapy in NSCLC with radiomics and pathomics, enhancing treatment

adaptation and personalization but at the cost of complex data integration processes. Deep learning techniques have

been prominently featured in several studies. Chen et al.

with multimodality attention, significantly improving detection accuracy in PET/CT images and samples.

Nevertheless, the high computational requirements present a challenge for real-time scenarios. Mohamed and

Ezugwu

enhanced lung cancer classification and prediction through deep learning and multiple omics data, achieving

high accuracy but grappling with the high dimensionality of the data samples.

On the analytical front, Qureshi et al.

visualized protein‒drug interactions to analyze drug resistance in lung cancer,

providing insights into key interactions but limited them to specific protein‒drug data samples. Alzubaidi et al.

devised a framework for lung cancer detection using CT scan images, employing both global and local feature

parameter that balances the contributions of both feature maps. This operation ensures that both local and global

features are effectively utilized for the final classification task. The final step involves a fully connected layer that

detection of small NSCLC nodules, which are often difficult to detect because of the inherent resolution limitations of

traditional imaging techniques. The SRGAN architecture leverages the adversarial training framework, which consists

of a generator and a discriminator network, to produce high-quality super-resolution images and samples. The

generator network in the SRGAN is responsible for upscaling the low-resolution input images and samples. Given a

for different scenarios. The generator network employs a series of convolutional layers, batch normalization, and

parametric rectified linear unit (PReLU) activations to progressively refine the image details. Mathematically, the

generator is described by Equation 5:

The integration of the SRGAN with other deep learning techniques, such as the CNN-Transformer hybrid model for

thereby improving the overall performance of the detection pipeline. The combination of these methods addresses both

the resolution and feature extraction challenges, leading to a more robust and accurate system for detecting and

classifying small NSCLC nodules. This residual connection helps preserve the original information while enhancing

further processing and final classification. The choice of the self-attention mechanism is justified by its ability to

dynamically adjust the focus on different parts of the image on the basis of their relevance, which is crucial for detecting

subtle features indicative of small NSCLC nodules. By emphasizing important regions, the self-attention network

complements the CNN-Transformer hybrid and the SRGAN by providing a mechanism to enhance feature extraction

and improve the overall model performance. This complementary nature ensures that the strengths of each component

are effectively utilized, leading to a robust and accurate detection and classification system.

the proposed classification process is shown in Fig. 2.

To ensure robustness and generalizability, the dataset is split into training, validation, and test sets at an 80:10:10 ratio.

The dataset used in this study is sourced from publicly available repositories provided by the National Cancer Institute,

along with clinical imaging data obtained from Acharya Vinoba Bhave Rural Hospital, Wardha, Maharashtra. The

training set includes 1,600 CT scans, the validation set comprises 200 CT scans, and the test set contains 200 CT scans.

Each subset is balanced to ensure a representative distribution of non-small cell lung cancer (NSCLC) nodules of

varying sizes, including those smaller than 3 mm. All data samples are fully anonymized prior to access, ensuring that

no personally identifiable information is available to the researchers. This study is based on fully anonymized (de-

The experimental setup for this study is designed as a comprehensive pipeline for evaluating the proposed hybrid deep

learning framework for detecting and classifying NSCLC nodules smaller than 3 mm, as well as predicting recurrence

risk. It includes key stages such as data acquisition, preprocessing, model training, validation, and performance

evaluation. High-resolution CT scans are utilized to capture fine-grained features essential for accurate diagnosis. The

framework leverages advanced deep learning techniques along with robust evaluation metrics to ensure reliable and

effective performance assessment. This experimental design enables thorough validation of the proposed approach and

demonstrates its applicability in real-world clinical settings.

The performance of the proposed framework is evaluated using several metrics. For the detection and classification of

4.3 Comparative analysis

The results were compared against those of three existing methods, represented as [13], [22], and [26] in Table 2. The

evaluation metrics included the accuracy, precision, recall, F1 score, and AUC-ROC for nodule detection and

classification and the mean squared error (MSE), mean absolute error (MAE), and R-squared (R²) score for recurrence

risk prediction. Additionally, interpretability scores were assessed using layer-wise relevance propagation (LRP).

The proposed model achieved the highest accuracy of 95.0%, significantly outperforming the other methods. Method

[26] achieved the closest performance (90.1%), indicating the effectiveness of the CNN-Transformer hybrid model in

detecting small NSCLC nodules. The detailed quantitative results are presented in Table 2, while the comparative

performance is illustrated in Fig. 3. As observed, the proposed model consistently achieves superior performance

In terms of nodule classification, the proposed model demonstrated superior performance across all the metrics. It

achieved a precision of 93.5%, a recall of 92.0%, an F1 score of 92.7%, and an AUC-ROC of 94.0%, highlighting the

robustness of the integrated self-attention mechanisms.

As summarized in Table 3, the SRGAN component of the proposed model significantly improved image quality,

achieving a peak signal-to-noise ratio (PSNR) of 32.5 dB and a structural similarity index (SSIM) of 0.92. This

improvement in image quality is crucial for the accurate detection and classification of small nodules.

As shown in Table 4, for recurrence risk prediction, the proposed LSTM network achieved an MSE of 0.030, an MAE

of 0.150, and an R² score of 0.90, demonstrating its superior ability to capture temporal patterns and predict the risk of

recurrence accurately.