Transfer Learning with Pre-Trained CNN Models: Advancing Colorectal Cancer Diagnosis

Histopathological image analysis is vital for timely colorectal cancer diagnosis. Our study uses transfer learning techniques to create an efficient and robust classification method. Within the present study, we utilize deep-learning models to develop a diagnostic system for colorectal cancer. Specifically, we employ pre-trained models, Inception V3, Xception, and MobileNet, to extract significant features from colorectal cancer histopathological images. Our experiment uses the CRC-VAL-HE-7K dataset with 7,180 histopathological images in nine classes, but we concentrate on a substantial subset of 5,000 images from six key classes. We train and evaluate deep transfer learning models utilizing the Adam optimization algorithm and categorical cross-entropy loss on this colorectal cancer histopathological image dataset. We assess the performance of these transfer learning models using key metrics: accuracy, sensitivity, precision, specificity, and F1-score. The results reveal that the Inception V3, Xception, and MobileNet models have outstanding accuracy during training, validation, and testing phases, with accuracy rates of 99.79%, 97.67%, 97.70%, 99.47%, 98.83%, and 98.40%, as well as 99.85%, 99.00%, and 98.30%, respectively. Our study presents an approach to colorectal cancer classification, leveraging deep learning techniques. The results have significant implications for early cancer diagnosis, offering a promising avenue for future medical applications.