Enhanced Malignancy Detection in Lung and Colon Histopathology Images Using Transfer Learning and Class-Specific Image Processing

Abstract

This study presents a robust framework for malignancy detection in lung and colon histopathology images, integrating class-specific preprocessing techniques with advanced transfer learning models. Leveraging the LC25000 dataset, the framework employs EfficientNetB0, ResNet-50, and InceptionV3 architectures to classify benign and malignant tissue samples. Tailored preprocessing methods, such as histogram equalization for lung images and edge detection for colon images, enhance feature visibility, enabling the models to focus on diagnostically relevant patterns. Among the models tested, EfficientNetB0 achieved the highest performance, with an accuracy of 95.3%, precision of 95.8%, recall of 95.0%, F1-score of 95.4%, and a ROC-AUC of 0.98. These results highlight the framework's effectiveness in balancing sensitivity and specificity, critical for clinical applications. Confusion matrix analysis further demonstrated EfficientNetB0's reliability, with minimal false positives and false negatives. However, addressing the false negative cases remains a priority to mitigate the risk of missed cancer diagnoses. While the framework is primarily validated on the LC25000 dataset, future work will incorporate additional datasets to enhance its generalizability. Furthermore, integrating ensemble techniques and advanced interpretability tools like SHAP and Grad-CAM will improve performance and clinical trust. This framework demonstrates significant potential for automated histopathology analysis, offering an accurate, interpretable, and scalable solution for malignancy detection in clinical workflows. By bridging the gap between AI advancements and medical diagnostics, it contributes to early and reliable cancer detection, ultimately aiding in better patient outcomes.