Non-Small Cell Lung Cancer Recurrence Prediction from CT Scans Using a Hybrid CNN-Vision Transformer Model

Artificial Intelligence (AI) in analyzing computed tomography (CT) images has significantly advanced computerassisted diagnosis (CAD) systems for lung cancer (LC). Nevertheless, there are complications like the complicated anatomy, superimposition of radiographic features and mobile positions of abnormalities which complicate obtaining the features to be used in successful CAD implementation. The cause of lung cancer is a major cause of disability in the entire world. The initial diagnosis is vital in preventing the progression and the development of effective management strategies. In this paper, the author suggests Hybrid-RViT, a hybrid neural network that combines both the local texture capture of convolutional neural networks (CNNs) and limited global context modeling capability of Vision Transformers (ViTs) to predict non-small cell lung cancer (NSCLC) recurrence with high accuracy on CT scans. The framework also presents a number of innovations: Smoothbased data augmentation (Gaussian blur, median filtering, bilateral smoothing) to increase the size of the dataset to 6,000 to augment the representation in transformers; CLAHE-based contrast enhancement to enhance the visibility of noisy images; and Intensity-Based Patch Tokenization (IBPT), which divides the CT images into low-, mid-, and high-intensity regions to increase the representation in transformers to 12 channels. CNN and ViT branch features are fused by Multi-Head Cross-Attention (MHCA) mechanism and they are classified by a ResMLP head. The Hybrid-RViT using ResNet50 obtained an accuracy of 97.12 percent and a macro-F1 of 0.97 and ROC-AUC of 0.9982 which is a$\mathbf{1 9. 2 8}$percent absolute and$\mathbf{2 4. 7 7}$relative increase over the baseline Balanced ViT. To make it clinical, Explainable AI (XAI) techniques were used, namely, Grad-CAM, SHAP, LIME, used to make it interpretable. The trained model was quantized using TensorFlow Lite to run on real time as a light weight, transparent and efficient diagnostic system that could be used on early lung cancer screening on a low resource clinical setting.former (ViT), Convolutional Neural Network (CNN), CrossAttention Fusion, Explainable AI (XAI), CT Scan