Deep Learning-Based Histopathological Analysis for Colon Cancer Diagnosis: A Comparative Study of CNN and Transformer Models with Image Preprocessing Techniques

Colon cancer is a serious life-threatening type of cancer that affects the colon and rectum. Manual diagnosis of this disease is time-consuming and la-bor-intensive, requiring pathologists to carefully examine multiple tissue sections and slides per patient. Deep learning techniques offer promising solutions to these challenges. This research investigates the performance of various deep learning architectures, including ResNet50, InceptionResnetV2, NasNet, EfficientNetB0, DenseNet169, ViT, and MobileViT, in histopathology. Four stain normalization methods (Vahadane, Macenko, Reinhard, and ITK) and four image-denoising techniques (Wavelet-based denoising, Gaussian Filtering, Bilateral Filtering, and Median Filtering) were employed to ad-dress staining and noise issues. An ensemble is developed integrating a CNN and a transformer. CNNs excel at extracting local features from images through the use of convolutional layers and pooling operations, while the transformer can capture long-range dependencies. This allows the ensemble to capture intricate local patterns while also considering the relationships be-tween objects, regions, or features that may span the entire image. A quantitative comparative study on the ‘EBHI’ dataset was conducted, resulting in the development of an ensemble model achieving an accuracy of up to 92.02%.