Innovative Deep Learning Strategies: EfficientNet and Dense Networks for Histopathological Colon Cancer Diagnosis

A common, serious disease that affects the colon is colon cancer.Since precancerous polyps frequently give rise to colon cancer, early detection and management are necessary for both effective treatment and higher survival rates.This study proposes a hybrid deep learning (DL) model to improve the detection of colon cancer from histopathological images, combining an EfficientNetB3 and Deep Dense Neural Network (DDNN).The study focuses on a balanced classification of tissue samples, consisting of 500 images of colon tissue having 250 normal and 250 adenocarcinomas utilizing the LC25000 dataset, having 25,000 histopathological images.The dataset was augmented to ensure comprehensive training and to reduce bias.The proposed model attained a remarkable accuracy of 98.00%, having precision of 97% for adenocarcinoma and 98% for normal tissue.A high recall value demonstrates the model's effectiveness in capturing real positive cases.The model achieved 98% F1 score for both classes, representing a successful balance between precision and recall.The area under the ROC curve (AUC) was 99.7% and five-fold crossvalidation yielded a consistent mean accuracy of 97.9% (0.25%).Repeated runs with different random initializations yielded consistent results, with an average accuracy of 98.0% (0.12%, 95% CI: 97.85-98.15).External validation using the NCT-CRC-HE-100K dataset, which includes 100,000 histopathological images from 86 patients, confirmed the model's generalizability with 97.3% accuracy.Computational efficiency analysis revealed that the hybrid model justified the trade-off for superior performance.This hybrid approach significantly advances colon cancer detection in histopathology images by leveraging sophisticated deep learning architectures, demonstrating high reliability, generalizability, and clinical potential.