TR-ABDOMEN-RAD — YOLO26-S-P2 demo weights

Six-class acute abdominal emergency detection on CT. Demo model trained on a balanced, patient-disjoint subset of TR_ABDOMEN_RAD — 6.8% of the positive slices (2.3% of all slices), 10 epochs, one 16 GB GPU.

Research/demo only. Not a medical device and not for clinical use.

Results (patient level)

Metric	Val (36 patients)	Test (26 patients)
Macro AUROC	0.879 [0.808–0.935]	0.921 [0.867–0.967]
Macro F1	0.781	0.664
Precision	0.763	0.644
Recall	0.806	0.733
Localization (restricted)	0.889	0.870

Slice-level detection: mAP50 0.220 / mAP50-95 0.092 (val). Weakest class: diverticulitis (AUROC 0.692); strongest: AAA (1.000).

Training

YOLO26-S-P2, 9.67 M parameters, 1280 px, bf16
10 epochs, batch 4 (effective 64), AdamW 7e-4, cosine schedule
Mosaic for the first 6/10 epochs (close-mosaic in the last 4)
10,764 train / 3,148 val / 2,828 test images

Files

best.pt                  final model (20 MB)
last.pt                  final-epoch weights
checkpoints/epoch*.pt    full training checkpoints (10 × 75 MB)
smoke/                   smoke-test weights (mAP 0 — ignore)
pretrained/              upstream YOLO weights used for initialization
results.csv              training log
eval_report.json         patient-level evaluation report
eval/scores_*.csv        per-patient scores
pred/                    raw val/test detections
plots/                   training curves, confusion matrix, PR
demo/                    GT-vs-prediction overlays

Limitations

6.8% of the positive slices (2.3% of all slices) and 10 epochs: a demo, not a paper reproduction.
26 test patients → wide confidence intervals.
Specificity was not evaluated.

Classes

acute appendicitis · acute cholecystitis · acute pancreatitis · kidney/ureteral stone · acute diverticulitis · AAA/dissection

Dataset

TR_ABDOMEN_RAD (introduced at TEKNOFEST-2022). Our corrected labels contain 1,296 studies / 43,166 boxes; the published Table 1 reports 1,274 patients / 42,922 boxes (896/189/189) — the +22 studies are mostly stones/AAA. Dataset terms apply.

References

Method (baseline): Erdoğan, H.M., Koç, U. Anatomically Localized Detection of Six Acute Abdominal Emergencies on CT Using Multi-window Deep Learning: Development and Validation. Journal of Imaging Informatics in Medicine (2026). https://doi.org/10.1007/s10278-026-02084-x

Dataset: Koç, U., Akçapınar Sezer, E., Özkaya, Y.A., et al. Elevating healthcare through artificial intelligence: analyzing the abdominal emergencies data set (TR_ABDOMEN_RAD_EMERGENCY) at TEKNOFEST-2022. European Radiology 34, 3588–3597 (2024). https://doi.org/10.1007/s00330-023-10391-y

Downloads last month: 1,239