Computer Aid-System to Identify the First Stage of Prostate Cancer Through Deep-Learning Techniques
José Gabriel García, Adrián Colomer, Fernando López-Mir, José M. Mossi, Valery Naranjo · 2019
Nowadays, there are high rates of discordance between pathologists when they analyse the biopsy samples to diagnose prostate cancer according to the Gleason scale. Thus, we designed a computer-aid system capable of accurately differentiating between normal tissues and pathological ones at the first stage. Specifically, we made use of an original segmentation algorithm to identify regions of interest and distinguish from them between artefacts (false glands), benign glands and Gleason grade 3 glands. Regarding the building of predictive models, we applied, for the first time, deep-learning algorithms on the previously segmented gland candidates. We compared the results reported by two different convolutional neural networks (CNNs) addressed with distinct classification strategies. The best model reached a multi-class classification accuracy of 0.812±0.033, after performing an in-depth data partitioning per medical history.