A 3D+2D CNN Approach Incorporating Boundary Loss for Stroke Lesion Segmentation

Dice loss is the most widely used loss function in deep learning methods for unbalanced medical image segmentation. The main limitation of Dice loss is that it weighs different parts of the to-be-segmented region of interest (ROI) equally, which is inappropriate given that the fuzzy boundary is typically more challenging to segment than central parts. A recently-proposed boundary loss weighs different parts of an ROI according to their distances to the ROI’s boundary, thus providing complementary information to Dice loss. However, boundary loss can not be directly applied to patch-based 3D convolutional neural networks (CNNs), significantly limiting its utility. In this paper, we proposed and validated a two-stage 3D+2D framework making use of 3D CNN for spatial information extraction and also boundary loss to complement the typically-used generalized Dice loss, for segmenting stroke lesions from magnetic resonance (MR) images. A 3D patch-based fully convolutional network was firstly used to learn local spatial features. And then the to-be-segmented MR image and the probability map predicted from the trained 3D model were sliced and fed into a 2D network with a joint loss combining boundary loss and generalized Dice loss. We evaluated the proposed method on a publicly-available dataset consisting of 229 T1-weighted MR images. The proposed approach yielded an average Dice score of 56.25% and an average Hausdorff distance of 27.14 mm, performing much better than existing state-of-the-art stroke lesion segmentation methods.