A Comprehensive Study of CNN Models for Colon Cancer Detection

Recent advancements in deep learning have significantly improved colon cancer detection through histopathological image analysis. CNNs such as DenseNet, MobileNet, ResNet, and GoogLeNet demonstrate strong performance, with each excelling in different aspects. DenseNet and ResNet enable deep feature extraction, while MobileNet offers lightweight, efficient deployment for low-resource clinical settings. This study presents a comparative evaluation of these models on colon cancer datasets. MobileNetV3 achieved the highest accuracy of 88.5%, followed by DenseNet121 (87.8%), ResNet50 (87.5%), and GoogLeNet (63.5%). Optimization strategies like data augmentation, transfer learning, and ensemble methods are also explored to enhance real-world applicability. A key contribution of this study is the novel combination and evaluation of DenseNet121, MobileNetV3, ResNet50, and GoogLeNet on the Kvasir dataset, which has not been collectively analyzed in prior literature for colon cancer detection. To further enhance their applicability in real-world scenarios, optimization strategies such as data augmentation, transfer learning, and ensemble approaches are also explored and discussed. A key contribution of this work lies in the novel combined evaluation of DenseNet121, MobileNetV3, ResNet50, and GoogLeNet on the Kvasir dataset. To the best of our knowledge, this is the first study that collectively analyzes these four models for colon cancer detection, thereby providing a comprehensive benchmark for future research in this domain.