An automatic multi-view disease detection system via Collective Deep Region-based Feature Representation

摘要

With today's growing requirements in disease diagnosis, we are constantly looking for better solutions. To meet the current demands, a disease detection system being highly effective as well as efficient is required. Existing and popular medical biometrics methods mainly focus on the local features extracted from raw medical image data, rather than study them globally. Meanwhile, prior knowledge is predefined in these methods so that procedures are inconsistent and require more manual operations. To address these, we present an automatic multi-view disease detection system, which contains a series of automatic procedures. The system first takes a tuple of images containing the face, tongue, and sublingual vein as the multi-view input, before directly outputting the predicted class label. To perform multi-view disease diagnosis, we propose a collective deep region-based feature representation. In summary, there are three real innovations in this paper: (1) Automated end-to-end medical biometrics system, (2) Deep region-based feature representation, (3) Multi-view multi-disease medical biometrics diagnosis. Extensive experiments were conducted on four diseases and one healthy control group using binary classification, showing both the effectiveness and efficiency of the proposed system. The average accuracy achieved was 95.8%, 96.49%, 96%, and 96.8% for breast tumor, heart disease, fatty liver, and lung tumor versus healthy control group taking 0.0031s, 0.003s, 0.0046s, and 0.0033s to process each sample respectively.