3680 - Weakly-Supervised Brain Segmentation for MR Images

Purpose/Objective(s): Deep learning (DL) can automatically segment medical images via learning from annotated training sets. However, the large variety of MRI sequences pose a challenge in effectively constructing enough labeled data. Here, we show that DL models trained on T1WI suffers a significant drop in performance when applied on T2WI of the same structure or vice-versa. Hence the automatic segmentation of MRI requires manual annotations for various sequences to provide training data for DL models. This observation urges us to evaluate the effectiveness of weak annotation methods that is a faster way to generate training data.

Materials/Methods: We select brain segmentation for our study since it’s manual contouring is time consuming due to the large size of the structure and the clinically expected high accuracy of auto segmentation might be a challenge for a weak annotation method. The dataset consists of 50 T1WI and 50 T2WI with full brain masks. We split the data into two halves for training and testing. We compare full masks against two weak annotation methods: point- and slice-based. The point annotation includes a bounding box, and 10 randomly sampled points with inner and outer labels assigned to them for each axial slice. The slice annotation keeps every 5th slice inside the brain. The negative slices are also provided with background label. Both strategies have 5 times speed up relative to the full mask creation. In addition, mixing of full and weak annotation is also evaluated under the same annotation budget. For all the annotation methods, we train a 2D nnU-Net. It is a strong segmentation method with support of weak annotation. The point annotation is combined with progressive training in which we gradually expand the labels around the collected random points using pseudo masks. This step has no annotation cost and guide the model to learn a more accurate segmentation. We test on full segmentations and measure 3D dice score and the 95th percentile Hausdorff distance.

Results: The segmentation performances are in Table 1. The mixture of point and full annotation allows to label 42 images instead of 10 with full mask within the same annotation time and achieve similar performance to full mask training. While slice annotation can even surpass the performance of models trained on full mask annotation.

Conclusion: Our study shows that DL models can learn accurately segment images from weak annotations. Adding more weakly annotated training samples results in more accurate auto segmentation than full annotation using the same annotation time hence it has the potential to accelerate the labeling of multiple MRI sequences.