Computer Vision News - January 2017

Goal The purpose of the challenge was therefore to develop algorithms that can predict the tumor proliferation scores from whole slide images. The challenge included three tasks: prediction of the proliferation score based on mitosis counting; prediction of the proliferation score based on gene expression; mitosis detection. Mitosis counting Histological grading systems for invasive breast cancer provide useful indications about how bad the condition is: generally, well- differentiated tumors have better outcomes. The most widely used system is B&R (Bloom & Richardson), which comprises three components, one of which is mitotic activity, a strong indicator for invasive breast carcinoma expressed as the number of mitotic figures in a specific area of tissue. A score is attributed according to the density of mitotic figures (hyperchromatic objects devoid of a clear nuclear membrane) found in that area. Unfortunately, other objects can be also found which are similar in aspect but unrelated to the pathology, making the task more challenging for the human expert (see image below). Results A Koran team, Lunit Inc. , obtained the best results in all three tasks comprised in the challenge, without using additional data. They won by proposing a unified framework based on DCNN ( Deep convolutional neural networks ) and consisting of a fully automated system including three modules: an image processing component for handling whole slide images; a deep convolutional neural network for mitosis detection; and a proliferation scores prediction module. Their paper is here . Results were discussed at the TUPAC16 workshop at MICCAI , two months ago in Athens. The workshop presented talks by participants in the challenge, including the winning team. “Develop algorithms that can predict the tumor proliferation scores” Challenge Computer Vision News Challenge 31 Examples of apoptotic nuclei, most commonly mistaken as mitoses