AbdomenNet

A patient arrives with abdominal pain, and the guideline answer is contrast-enhanced CT or CT angiography. The scan that actually gets taken is often the non-contrast one — renal dysfunction, a contrast allergy, no venous access yet, or a saturated department. Non-contrast CT is precisely the study that lacks the vascular and perfusion signal making mesenteric ischemia or an aortic dissection legible, so the first image a radiologist sees is the one least equipped to answer the question being asked of it.

AbdomenNet, from Shanghai Sixth People's Hospital and the Shanghai Jiao Tong University School of Medicine, treats non-contrast abdominal CT as a domain worth its own foundation model rather than as degraded contrast CT. A DINOv2 self-distillation objective is run over 103,989 unlabeled abdominal non-contrast examinations from 66,975 patients, and the resulting backbone is fine-tuned once into a multi-task head covering 11 acute abdominal conditions plus three surgical risk-stratification subtasks, then evaluated unchanged at three independent hospitals. (An unrelated 2022 network for abdominal organ segmentation on Dixon MRI carries the same name.)

The headline finding is about pretraining provenance rather than scale. Image-only pretraining on in-domain non-contrast scans beat every alternative tried, including report-paired vision-language pretraining on the same corpus — emergency reports are too brief and too contaminated by later contrast-enhanced findings to supervise signals that must be caught before any report exists.

#Key Features

• Eleven conditions from one shared encoder: Aortic dissection, mesenteric ischemia, gastrointestinal perforation, intestinal obstruction, visceral injury, cholecystitis, choledocholithiasis, pancreatitis, appendicitis, diverticulitis and ureteral stone are predicted jointly, with labels treated as non-exclusive.
• Severity, not only presence: Three cascaded subtasks grade high-grade visceral injury (AAST grade ≥III), strangulated obstruction and upper gastrointestinal perforation — the distinctions that decide whether a patient goes to surgery.
• Volumetric context from a 2D backbone: Each axial slice is encoded independently, then a three-layer Transformer decoder aggregates the sequence into a scan-level prediction, recovering inter-slice dependencies a linear head would discard.
• Calibrated for missing nothing: Triage thresholds are fixed on the validation set at sensitivity ≥0.95 for the five emergent conditions and two risk subtasks, accepting false positives as the price of a low miss rate.
• Attention checked against anatomy: Grad-CAM maps are quantified against TotalSegmentator-derived organ regions rather than assessed by eye.

#Technical Details

The backbone is ViT-g/14 — 1.1 billion parameters, 14×14 patches — pretrained by DINOv2 on 2D axial slices at 518×518 resolution, with volumes windowed to soft tissue (width 350 HU, level 40 HU) in both stages. Fine-tuning used 5,816 annotated patients from the lead center. On the internal test set the macro-average AUROC across the 11 conditions was 0.908, rising to 0.937 for the five emergent conditions; across 2,528 patients at three external centers it held at 0.898, 0.889 and 0.896, with 0.919 for emergent conditions. Ablations isolate the contribution of domain pretraining: training from scratch reached 0.714, general-domain DINOv2 initialization 0.822, SimCLR 0.794 and MAE 0.750 on the same corpus. Under a harmonized protocol, Merlin scored 0.678 with its released weights and 0.833 re-pretrained on this corpus with paired reports, CT-FM 0.823, TAP-CT 0.845 and Curia 0.874.

#Applications

In a four-radiologist crossover reader study over 380 cases with a three-month washout, AI assistance raised senior readers' AUROC on emergent conditions from 0.876 to 0.968 and junior readers' from 0.735 to 0.874, narrowing the experience gap, while cutting median reading time from 197.5 to 145.0 seconds. A retrospective reconstruction over 18,398 emergency CT examinations re-ordered the reporting queue by model output: median turnaround for emergent cases fell from 49.3 to 12.3 minutes with no significant delay to the remaining queue. The intended role is escalation triage — flagging which non-contrast studies warrant urgent contrast-enhanced imaging or surgical consultation — not replacing contrast CT.

#Impact

AbdomenNet is one of the few radiology foundation models evaluated end to end across generalization, reader benefit and workflow effect rather than on AUROC alone, and its ablations counterweight the assumption that paired reports always improve medical image pretraining. The caveats are substantial and the authors state them. Development and validation were entirely within Chinese tertiary centers, and the cohort was enriched for the target conditions — only about 22% of patients had none of them — so performance in unselected emergency populations will be lower. The high-sensitivity thresholds flagged 767 of 1,761 abdominal scans as top priority, of which 564 were false positives, and alert fatigue was not measured. The workflow gain is a reconstruction of historical queues, not a prospective trial. Code is Apache-2.0 and the self-supervised backbone checkpoint is distributed on HuggingFace behind manual access approval; the fine-tuned diagnostic heads are not released, and the patient imaging data is available only by application to the corresponding author.