Beyond Human Variability: Deep Learning for Intravascular Ultrasound Segmentation With Noisy Labels

Intravascular ultrasound (IVUS) is an imaging modality that provides cross-sectional visualization of the coronary lumen and vessels. To address the challenges of expertise and time-consuming annotation for IVUS interpretation, deep learning approaches have demonstrated accurate prediction in semantic segmentation. However, the impact of noisy labels from high inter-observer variability has not been evaluated systematically.

We analyzed the effect of noisy labels on deep learning-based IVUS segmentation and proposed a clinically informed filter to assess the quality of the generated labels, thereby improving the model performance in semi-supervised learning.

When noisy labels occurred in a consistent pattern or exhibited poor boundary alignment with the ground truth, segmentation performance deteriorated substantially. In contrast, when correct outlines were preserved in half of the boundaries, the degradation in the Dice similarity coefficient was limited to 1.92%, even with error amplitudes of 20 pixels. Increasing the training dataset size mitigated the adverse effects of the noisy labels. Furthermore, a filter based on the Hausdorff distance between the union and intersection of the predictions across three consecutive frames improved the segmentation performance in self-training.

These findings provide practical guidance for building training datasets that support the development of robust deep learning models for IVUS imaging.