2811

View Abstract / View Presentation

Deep Multiple Sclerosis Lesion Segmentation with Anatomical Convolution and Lesion-wise Loss

Hang Zhang¹, Jinwei Zhang¹, Pascal Spincemaille¹, Thanh D. Nguyen¹, and Yi Wang¹
¹Cornell University, New York, NY, United States

We propose an anatomical convolutional module and lesion-wise sphere loss for improving MS lesion segmentation.

Figure 1. Example Illustration of the proposed anatomical coordinates with one slice from the axial direction.

Figure 2. The computed anatomical coordinates are concatenated with the incoming feature tensor through channel dimension, followed by standard convolution, batch normalization, and ReLU activation for feature fusion. c is the number of channels in the feature tensor.