Supervisors: Dr. Ertunc Erdil, Prof. Ender Konukoglu
Quantifying segmentation uncertainty has become an important issue in the domain of medical image analysis. While some architectures have been developed to quantify the uncertainty of medical images, they mostly suffer from high memory requirements. We propose an architecture called ReversiblePHiSeg for image segmentation. ReversiblePHiSeg generates segmentation samples to quantify uncertainty and contains reversible blocks for memory efficiency. Reversible blocks reduce the required memory amount by not storing the activations in every layer but calculating them from the following layer during backpropagation. This leads to a smaller memory footprint with a slight increase in training time. This permits to train these neural networks on hardware with limited GPU memory and improve one of their major disadvantages. We demonstrate the advantage of the proposed method in various medical image segmentation problems.