Learning MRI and histology image mappings for cancer diagnosis and prognosis

点击翻译按钮获取中文摘要

This project aims to exploit recent advances in machine learning to address acute problems in cancer management - most directly prostate cancer. The current standard approach of making treatment decisions via biopsy and histology has two key limitations; it is invasive and subjective/inconsistent. We will develop the computational tools supporting new solutions that resolve both issues. Specifically, we aim to enable non-invasive MRI to become the primary diagnostic tool. This would avoid a large number of unnecessary biopsies, which carry significant risk of life-changing side-effects, reserving the procedure for only marginal cases. We also plan to relate MR signals to quantitative tissue features enabling consistent assessment and thus more reliable treatment decisions.The use of MRI in prostate cancer has become routine only in the last few years. Thus, data relating MRI to patient outcome (e.g. 5-10 year survival) is not available. However, we are uniquely positioned to obtain i) associated MRI and histology images, and ii) associated histology and patient outcome. In combination, these support a two-step learning and estimation process: from MRI to histological features; and from histological features to patient prognosis. Such mappings can provide invaluable new information for clinical decision making, as well as guide the design of maximally informative future MRI protocols. Such protocols will enable long-term data collection initiatives that support direct mappings from MRI to outcome.The project involves engineering challenges that demand innovations at the cutting edge of image-based machine learning technology: i) accommodating uncertainty in the alignment of training images; ii) quantification and visualization of uncertainty in the output of learned models; iii) salient feature selection in high-dimensional input data; iv) development of experiment design optimization algorithms driven by implicit computational models (such as neural networks). We build on the latest ideas in deep learning to address these challenges. We tailor solutions relevant to the immediate problems at hand in prostate cancer, but that extend to related tasks in cancer imaging and medical imaging in general.