Can deep-learning algorithms identify genetic mutations or aberrant cellular signalling pathways from medical images?

点击翻译按钮获取中文摘要

Our project will determine if cancer-related genetic mutations found in tumours can be detected using medicalimaging. Tumours arise from a series of genetic errors, and these determine much of the behaviour of a tumour,including its aggressiveness and its response to a treatment. In day-to-day medical imaging, once a patient hasundergone a medical scan, a specialist will look at the images and provide a diagnosis (e.g., liver cancer).Sometimes, a biopsy (tumour sample) is acquired to better determine the subtype of cancer. Our project aims atassisting physicians by providing them with additional information extracted using artificial intelligence andadvanced computer software. These are already known to be superior to humans in finding and quantifyingsubtle image characteristics. We hypothesize that these image characteristics could be predictive of the cancersubtype and its optimal treatment. First, this software has to be trained to recognize mutations. Because humantumours vary a lot, it is difficult to differentiate visual characteristics caused by an individual inherentvariability from those caused by the mutation. To overcome this, we will use genetically engineered mousemodels - these will have specific mutations that will result in cancer but with limited variability betweenanimals. This will allow us to train a software to recognize tumours that have specific mutations. If successful,our project will ultimately lead to software tools with capabilities similar to biopsies, and better and lessinvasive management of cancer.