Lung cancer severity detection with pulmonary pathology using optimized deep learning framework

Abstract

Lung cancer leads to mortality worldwide, and hence, accurate cancer grading is essential for prognosis and treatment planning. Existing tumor grading approaches are time-consuming and prone to inter-observer inconsistency. Also, existing Deep Learning (DL) approaches focus on binary classification and fail to identify the fine-grained tumor aggressiveness, which is crucial for accurate prognosis and treatment planning. This study develops a hybrid attention-based DL grading framework for accurate detection and grading of lung cancer into adenocarcinoma, squamous cell carcinoma, and normal tissues, which differentiates into grade 1, grade 2, and grade 3 tumor grades. The proposed methodology utilizes a Macenko color normalization method to preprocess the high-resolution histopathological images. The proposed work introduces a hybrid attention-based DL framework, ResCBAM-ViTFSG, that integrates ResNet18 with Convolutional Block Attention Module (CBAM) for extracting local spatial features and Vision Transformer with Feature Selection Gate (FSG) for extracting global features. To improve robustness, introduced a novel Uncertainty-Aware Reweighting Adam (UAR-Adam) optimizer that dynamically adjusts learning rates. The proposed work achieved fine-grained grading of lung cancer through per-class grading analysis, effective feature extraction counts, and optimizer benchmarking, achieving an accuracy of 98.34 %, precision of 98.50 %, recall of 98.18 %, F1-score of 98.34 %, and an inference time of 7.8 ms. These results outperform baseline models in lung cancer grading, suitable for deployment in pulmonary pathology clinical diagnostic systems.