Deep Learning Model for Classification and Segmentation of Tumor Regions in Histological Images with Emphasis on Interpretability

Breast cancer is one of the leading causes of cancer deaths in the world. Medical Image analysis field has been significantly developed in the last few years. It has been used in medical applications to diagnose pathologies, guide medical interventions or for research purpose. A lot of research has been done for the medical field however; few of them are used in real life application. There are several issues that affects the reliability of the models such as the bias in a small dataset; the variability of large internationally sourced datasets; the lack of labeled data for training, and the necessity for clinicians and analysts to work together to get the best AI result. In this project, we leverage multidisciplinary team with Computing and Health Science expertise to investigate a multi-resolution deep learning model for tumor classification and segmentation in breast histopathology images. The architecture of the model will be based on convolutional autoencoders, a type of neural network particularly tailored to learn a compressed representation of structured raw data such as images. An autoencoder is composed of an encoder and a decoder sub-models. The encoder can be used to perform feature extraction on raw images. The extracted features can then be used to train a different deep learning model to make patch-level predictions to discriminate tumor-patches from normal-patches. We generate labels for small image patches in order to infer locations of suspected abnormalities in a query image. To train the model we will use (1) public datasets, such as Camelyon16 Dataset provided by the organizers of the competition and the evaluation metrics used to rank the participants such as the area under the receiver operator (AUC) score, and (2) we will seek to collect data from Omani hospitals with focus on SQUH data.