Deep Learning and Features Fusion for Colorectal Cancer Detection from Histopathology Images
V. Rajinikanth, Ramya Mohan, Mathiyazhagan Narayanan
Abstract
Modern hospitals have extensively adopted the automatic disease examination methods to detect diseases from biomedical images of selected modalities. These methods reduce the diagnostic burden and enhance detection accuracy. This study seeks to create a Deep-Learning (DL) tool for the analysis of histopathology images related to colorectal cancer (CC). This DL-tool consists of the following stages: image collection and resizing, feature extraction using a selected DL-scheme, feature reduction through 50% dropout and serial feature fusion to create the Fused-Features-Vector (FFV), followed by binary classification and 3-fold cross-validation to validate the results. This study initially examines traditional DL-models for classification using SoftMax. Based on the results, the two best models are selected to generate the FFV. This study examines the features of VGG16 and ResNet101 to produce the FFV, thereby validating the effectiveness of the proposed scheme. The CC detection utilizing the FFV achieved an accuracy exceeding 95%, confirming that the developed scheme yields improved results on the selected data.