Sonderkolloquium, Jun.-Prof. Dr. Mattias Heinrich, Uni Lübeck / am 07.07.2017

07.07.2017 von 14:15 bis 15:45

Institut für Informatik,Christian-Albrechts-Platz 4, R.715, 24114 Kiel

Titel: Learning Sparse Binary Features for Medical Image Segmentation of the Abdomen

Abstract: In this talk, we explore the capabilities of sparse binary features for medical image segmentation. Due to insufficient contrast and anatomical shape variations local image patches rarely provide sufficient information for accurate segmentation of abdominal structures. Based on our two recent MICCAI papers, we propose to use long-range binary features to robustly capture the image context. Two different classification strategies are subsequently developed. 

First, a very fast approximate nearest neighbour search based on vantage point forests and Hamming distances between feature strings is presented. The classifier can be learned and applied to new data in few seconds. The approach reaches state-of-the-art performance for larger organs on the VISCERAL3 benchmark.

Second, we develop a deep neural network architecture that combines a local CNN path with a new contextual path that encodes the sparse binary features. Following the ideas from Network-in-Network, 1x1 convolutions are employed to learn the best combination of different binary offset locations. We demonstrate experimentally that this restricted feature extraction in the first layer enables to regularise the network with a huge receptive field and leads to short training times of less than 10 minutes. Using only 1 million trainable parameters, the model achieves a accuracy of 64.5% Dice, which is comparable to the best performing, much more complex deep CNN approach for pancreas segmentation.

Finally, the potential use of learned binary features for other tasks in medical image analysis, such as image registration and disease classification will be discussed.

 

Prof. Meyer

Diesen Termin meinem iCal-Kalender hinzufügen

zurück