Predicting Proliferation and Differentiation in Prostate Cancer : from H&E Images Using Deep Learning

Sammanfattning

Accurate assessment of prostate cancer, the most common malignancy among men, is needed toavoid overtreatment as many cases follow an indolent course. One such method is using immuno-histochemical staining to combine the biological information of Ki-67 for proliferation and PSA fordifferentiation - a process that is usually done by a professional pathologist through labor-intensiveand poorly standardized work. This report aims to automate and standardize immunohistochem-ical scoring using deep learning methods, as well as inquiring whether the same information canbe gathered from general tissue morphology using inexpensive H&E staining. A multitude of AI models were trained on a retrospectively gathered data set of 119 prostate cancerbiopsies containing Ki-67, PSA and H&E whole-slide images. Models were evaluated on attentionheat map - human observation coherence and through common model metrics to obtain the bestpossible model. Ki-67 score prediction from Ki-67 staining slides proved to be the best followed byPSA score prediction from PSA staining, Ki-67 scores from H&E slides and lastly PSA scores fromH&E. In conclusion the models were accurate in providing scores, and heat maps concentrated oncancer tissue while avoiding image artifacts. Predictions from H&E staining were accurate enoughto prove that general tissue morphology carries information of other biomarkers - larger and morediverse datasets are however needed for clinical validation.