Optimal feature extraction and ulcer classification from WCE image data using deep learning

Cancer is a difficult disease and one of the leading causes of human mortality. There are various types of cancers associated with various parts of human anatomy. Among all cancer-related deaths, gastrointestinal (GI) cancers like esophageal, stomach, and colorectal cancer have significantly higher mortality rates. There can be a great reduction in mortality if these pre-cancerous lesions are detected and removed in earlier stages. Manual methods for cancer detection are painful and time-consuming. Video has been used as a means of detection for this type of cancer. However, video having many frames requires a lot of human effort for better analysis and is prone to the wrong diagnosis at a greater level. Automatic intelligent detection and classification help in the reduction of the death rate among cancer patients. This also can assist experts in better diagnosis and treatment. This paper presents a novel framework for the automated classification of ulcers using the WCE image KVASIR dataset. First, preprocessing is done to enhance the contrast, which improves the classification task. Well-known deep learning models ResNet50 and ResNet152-V2 are trained using preprocessed data and trained features are extracted. These features are fused using concatenation. These combined features are optimized using modified GA. The optimal number of features is then used for classification and an accuracy of 99.67 has been achieved. Results are compared with current state-of-the-art techniques and the proposed method has performed better in terms of improved accuracy.