[2503.05916] SAS: Phase Something Small for Ultrasound — A Non-Generative Information Augmentation Approach for Strong Deep Studying in Ultrasound Imaging

View a PDF of the paper titled SAS: Phase Something Small for Ultrasound — A Non-Generative Information Augmentation Approach for Strong Deep Studying in Ultrasound Imaging, by Danielle L. Ferreira and 4 different authors

View PDF
HTML (experimental)

Summary:Correct segmentation of anatomical constructions in ultrasound (US) pictures, notably small ones, is difficult resulting from noise and variability in imaging circumstances (e.g., probe place, affected person anatomy, tissue traits and pathology). To handle this, we introduce Phase Something Small (SAS), a easy but efficient scale- and texture-aware information augmentation method designed to reinforce the efficiency of deep studying fashions for segmenting small anatomical constructions in ultrasound pictures. SAS employs a twin transformation technique: (1) simulating numerous organ scales by resizing and embedding organ thumbnails right into a black background, and (2) injecting noise into areas of curiosity to simulate various tissue textures. These transformations generate real looking and numerous coaching information with out introducing hallucinations or artifacts, enhancing the mannequin’s robustness to noise and variability. We fine-tuned a promptable basis mannequin on a managed organ-specific medical imaging dataset and evaluated its efficiency on one inside and 5 exterior datasets. Experimental outcomes display important enhancements in segmentation efficiency, with Cube rating positive aspects of as much as 0.35 and a median enchancment of 0.16 [95% CI 0.132,0.188]. Moreover, our iterative level prompts present exact management and adaptive refinement, reaching efficiency similar to bounding field prompts with simply two factors. SAS enhances mannequin robustness and generalizability throughout numerous anatomical constructions and imaging circumstances, notably for small constructions, with out compromising the accuracy of bigger ones. By providing a computationally environment friendly resolution that eliminates the necessity for in depth human labeling efforts, SAS emerges as a strong device for advancing medical picture evaluation, notably in resource-constrained settings.