Deep Learning Segmentation of Abdominal Organs Across Ultrasound, MRI, and CT: A Scoping Review of Modality-Specific...

Segmentation of the liver, kidneys, and pancreas underpins volumetry, treatment planning, and disease monitoring across ultrasound (US), magnetic resonance imaging (MRI), and computed tomography (CT). Deep learning has advanced through four waves-convolutional networks, transformers, segmentation foundation models, and large language model (LLM)-based agents-but how performance differs across organs and modalities has not been synthesized in a segmentation-centric, modality-stratified way. Following scoping-review reporting guidance, we searched five databases for studies published 2020-2026 and charted 72 studies by organ, modality, method family, dataset, and performance. Organ tractability was consistent (liver > kidney >  > pancreas), as was modality maturity (CT > MRI > US), with one benchmark-driven exception for the pancreas; tumor segmentation lagged organ segmentation throughout. Foundation models matched or exceeded equally box-prompted specialists on individual unseen tasks, but the most frequently cited result of this kind belongs to the Segment Anything Model (SAM) rather than to a medically adapted model, and every comparator in it received the same ground-truth-derived box. On the two benchmarks in our pool containing a fully automatic comparator, that comparator led on one and was within about one Dice point on the other, and no benchmark evaluated specialists, foundation models, and agents under a single common protocol. Performance degraded sharply on the pancreas and on ultrasound. LLM/agentic methods did not surpass specialists on overlap metrics, offering gains instead in generality, autonomy, and annotation efficiency. This modality-stratified synthesis shows where ultrasound and the pancreas remain underserved, and indicates that agentic approaches currently augment rather than replace specialist segmentation.