Show abstract - Show schedule - Proceedings - PDF - Reviews

Data augmentation techniques are frequently used to prevent overfitting, enhance generalizability, and overcome limited amounts of data. This data-oriented approach commonly includes domain-agnostic techniques of geometric transformations, colour space changes, and generative adversarial networks. However, leveraging domain-specific traits in aug- mentations may yield further improvements. We propose three new contributions to ultrasound augmentation: zoom, artificial shadowing, and speckle parameter maps. We first present zoom, a modification on scale which maintains the ultrasound beam shape. We then characterize acoustic shadows within abdominal ultrasound images, and formulate a method to introduce artificial shadows in a realistic manner into existing images. Finally, we transform B-mode ultrasound images into speckle parameter maps based on the Nakagami distribution to represent spatial structures not obvious in conventional B-mode. The three augmentations are evaluated in training a fully supervised network and a contrastive learning network for multi-class intra-organ semantic segmentation. Our preliminary results demonstrate the benefit of using zoom and speckle maps as augmentation, and the challenges presented by acoustic shadowing, in segmentation.
Hide abstract

Thursday 7th July
Poster Session 2.2 - onsite 11:00 - 12:00, virtual 15:20 - 16:20 (UTC+2)
Hide schedule