Segregation similarity loss in morphological ranking of image search in histopathology
MetadataShow full item record
Histopathology is the study of changes in tissue caused by diseases such as cancer. It plays an important role to diagnose the cancers. Regarding the large variation of many cancers types, and the large size of Whole Slide Images (WSIs), the analysis of histopathology images is challenging. To come up with this challenge, AI algorithms, such as deep learning (DL) are used to automate image analysis efficiently and accurately. In this study, some DL methods are developed on medical images focusing on the following goals. Firstly, we propose a model that can help us classify cancers better and faster and achieve good results compared to other models. Most of the present models for the classification of histopathology images are very large and accordingly have many parameters to be learned/optimized and require enormous computational times to achieve reliable results. We propose a more compact network that is tuned to classify cancer subtypes with less computation time and memory complexity to overcome these issues. This model, namely custom EfficientNet, is based on EfficientNet topology, but it is tailored for classifying histopathology images. The utilized model is evaluated over three-tumor-type brain, lung, and kidney from TCGA repository. The results show that the proposed model, compared to state-of-the-art models, i.e., KimiaNet, can classify cancer subtypes more accurately and provides superior results. Besides, the proposed model achieves memory and computational efficiency in the training phase and is a more compact deep topology compared to KimiaNet. More recently, deep learning was applied for the challenging task of image search on the TCGA repository. Researchers can use the image search results to compare data of current and previous patients and learn from cases that have been clearly treated and diagnosed. However, there is no way to train a model using image search, hence image search must be used on the outcomes of a model that was trained using classification. Moreover, it can be seen that the obtained results from the classification method suffer from bias, and the classification loss function cannot make it possible for us to reduce bias during the network’s training phase. Secondly, the research proposes a new loss function, Similarity Loss (SL), to address these problems. This loss function allows us to train the model based on image search, removing the requirement for us to use other approaches for training image search models. Besides, unlike the classification loss function, the modified version of this loss function, Segregation Similarity Loss (SSL), helps us reduce the adverse effect of one of the major problems in this field called bias and obtain better and more reliable results. By utilizing SSL, we achieve promising results to classify histopathology images. SSL function achieved up to 5% and 9% improvement compared with the state-of-art models for Lung and Brain dataset, respectively.