- [pdf] [supp]
SUPRA: Superpixel Guided Loss for Improved Multi-Modal Segmentation in Endoscopy
Domain shift is a well-known problem in the medical imaging community. In particular, for endoscopic image analysis data can have different modalities that cause the performance of deep learning (DL) methods to become adversely affected. Methods developed on one modality cannot be used for a different modality without retraining. However, in real clinical settings, endoscopists switch between modalities depending on the specifics of the condition being explored. In this paper, we explore domain generalisation to enable DL methods to be used in such scenarios. To this extent, we propose to use superpixels generated with Simple Linear Iterative Clustering (SLIC), which we refer to as "SUPRA" for SUPeRpixel Augmented method. SUPRA first generates a preliminary segmentation mask making use of our new loss "SLICLoss" that encourages both an accurate and superpixel-consistent segmentation. We demonstrate that SLICLoss when combined with Binary Cross Entropy loss (BCE) can improve the model's generalisability with data that presents significant domain shift due to a change in lighting modalities. We validate this novel compound loss on a vanilla UNet using the EndoUDA dataset, which contains images for Barret's Esophagus from two modalities. We show that our method yields a relative improvement of more than 20% IoU in the target domain set compared to the baseline.