63 — Adversarial Networks for the Detection of Aggressive Prostate Cancer

This project explores the challenge of image-segmentation of medical images to detect aggressive types of prostate cancer. Using 3T MRI imagery, Kohl et al use fully-convolutional neural networks (FCNs) to semantically segment images of prostate, looking for particularly aggressive cancerous growth.

Compared to other CNNs, this network uses an adversarial architecture to target and interrogate subsections of the image that have particularly high candidacy for cancer-detection. The GAN is trained on a subpopulation of the dataset with data duplication (shift and rotation) to increase the size of the training set.

The results of this work suggest that the combination of semantic segmentation — where class of tissue (tumor, peripheral zone, transition zone) is encoded in the resulting segmentation — and generative-adversarial models (GANs) is particularly potent for detecting tumors in prostate tissue.

I hope that this work inspires more combination of GAN/semantics in medical imagery, where the generative component helps to increase the size of the training dataset (which is commonly very small in medical study) and the semantic segmentation aspect assists in increasing the specificity of the resulting segmentation.