A Generalized Deep Learning Framework for Enhancing Colorectal Cancer Diagnosis Through Transfer Learning and Cross-Dataset Validation

Colorectal cancer (CRC) is a leading cause of cancer related death across the globe, with accurate identification at the early stage of malignancy to be of reduced significance to improve treatment outcomes. However, automated diagnostic systems tend to suffer from a lack of reliability from the variability of the datasets, domain shifts, and a limited generalization of heterogeneous medical data imaging sources. To solve these issues, a multi-stage deep learning model for predicting colorectal cancer from images of histopathology specimens is proposed in this paper. The framework combines the M-Net used for Segmentation of Tumors (U-Net), EfficientNet used for extracting features of the image (Deep Learning)(DL), and DenseNet with Transfer Learning used for precise classification of the cancerous and noncancerous tissues. The proposed approach includes the use of a structured pre-processing, data augmentation and cross-catalog validation to improve robustness and minimize distribution mismatch of heterogeneous datasets. Extensive experiments using several colorectal cancer data sets, such as NCT-CRC-HE-100K, CRC-HGD-v1, CRCCDV1 and the Cancer Genome Atlas (TCGA) show the efficacy of the proposed framework. The model has up to 96.5 % accuracy, 95.2 % sensitivity and 97.2 % specificity with an average cross-dataset improvement of 8.7 % and 35 % reduction in training time compared with conventional deep learning models. The results of the research underscore the ability of the proposed framework to enhance the reliability, generalisability and computational effectiveness of diagnostic work for colorectal cancer supporting the development of robust and clinically deployable colorectal cancer diagnostic systems in varied imaging environments.