43007 Tarragona, Spain
Digital histopathology is a central part of the prediction workflow, for cancer prediction, tumour grading and planning of suitable, personalized treatment strategies. The major bottleneck in this process however, is the huge volume of data contained within digitized histology slides (for a single patient) and the subjective and time consuming of analysing them, undertaken by expert pathologists. Consequently, automatic systems for detection and prediction have a crucial role to play in the future, by supplementing the diagnostic workflow and aiding pathologists such that their efforts may be concentrated on the more vital aspects of diagnosis and treatment planning. With this in mind, the main objectives of this project are: (1) Formulate an end-to-end DL model that enables the analysis of HE-stained WSIs images of BC specimens, for the joint task of detection and identification of malignancy and BC subtypes; (2) Investigate the benefits of learning features across multiple scales/magnifications, in order to improve the detection and prediction of cancer; (3) Establish a multitask framework that incorporates related auxiliary tasks such as mitosis detection, cellularity assessment and nuclei segmentation, for improved prediction and cancer staging/tumour grading; and (4) Evaluate the benefits of simultaneously learning from both HE stained and immunohistochemically (IHC) stained WSIs, for the identification of the molecular sub-types of BC and relapse prediction.
As a result, the outcomes of this project are: (1) a framework that is trained on sparse labels (e.g., presence or absence of cancer) to predict the risk of cancer within an unseen WSI and provide a heat map of regions likely to contain cancerous cells; (2) Learning features across multiple scales is expected to ultimately lower the workload of pathologists as accurate prediction of suspicious regions in low magnification images enabling pathologists to concentrate on a small subset of the data, at high magnifications; (3) Incorporating auxiliary tasks into the learning framework is expected to enable further categorization of how aggressive the cancer is, i.e. cancer staging, which is essential for formulating suitable and personalized treatment plans, and for the overall prognosis of the patient; (4) Jointly learning from HE and IHC-stained WSIs, is expected to yield a general set of features such that molecular sub-types of BC can be identified directly from the HE-stained WSIs alone or to improve the capability of cancer relapse prediction.