Deep Learning for Assessing Risk of Prostate Cancer

Cancer detection is one of the leading research topics in medical science. Prostate cancer is the second most common cancer in men and the sixth leading cause of cancer death among men in the world. The current diagnostic pathway is based on serum Prostate-Specific Antigen (PSA) levels. Although PSA screening reduces the spread and death from cancer, it overdiagnoses some low-risk cancers that may not have caused damage, leading to unnecessary invasive examinations. Several methods have been proposed in the past, such as studying the evolution of prostate antigen over time using different velocity formulas, which have often led to inconsistent results. In this thesis, some Deep Learning methods have been applied to time series to predict prostate cancer risk. The main objectives are early diagnosis and reduction of the number of unnecessary biopsies. Several approaches have been proposed to deal with the irregularity of the time series, one of the most recurrent problems in clinical data. The first technique is based on regularizing the time series, while the second adds new features, such as the time distance between visits. The results have been compared using different metrics and finally, the best model is selected. The performance obtained suggests that the proposed methods are promising and can be a helpful tool to support clinical decision-making.