43007 Tarragona, Spain
The main objective of this project is to develop and train transparent DL models for analysing radiological data for diagnostic purposes in BC and to recognize anatomic signs of predisposition for BC. Specifically, we will mainly work with tomosynthesis (3D mammography) and 2D mammography images. In particular, the following tasks will be performed: (1) Detect suspicious masses in tomosynthesis and mammograms with DL. (2) Automatically Identify of the most aggressive BC subtypes with hand-crafted and DL methods. (3) Develop a DL model capable of estimating the uncertainty of the predictions in order to assess the benefit from additional MRI/ultrasound screening. (4) Develop an explainable-DL model capable of predicting the long-term risk of relapse after treatment and the short-term responsiveness to non-surgical treatments. (5) Study the correlation between the common molecular subtypes and the changes in radiological images of BC patients after treatment. (6) Combine the developed methods into a clinical workflow designed to maximize patient outcome and allocation of available resources. In order to perform these tasks, we have access to one of the biggest datasets on tomosynthesis in the world through our close collaboration with SUH providing a large dataset of mammograms.
We will provide transparent DL models on radiological images that can be able to (1) detect cancer and its subtype, (2) estimate the uncertainty in the estimations, (2) estimate the risk of relapse and responsiveness to treatment, (3) follow-up the evolution of tumours with longitudinal data. These models will be used to design an improved clinical workflow for BC. This workflow will use better estimates of risk to help reduce mortality by screening high-risk patients with more sensitive MRI scans. The outcomes will be based on accessing large amounts of training data for developing trustworthy DL methods in order to push further the state-of-the-art. Moreover, the manual analysis of tomosynthesis images is a big burden for radiologists due to its dimensionality (3D) compared to 2D mammograms. Thus, in this project we will assess the power of AI-based automatic tools for the analysis of tomosynthesis.