skblv/yolo11m-cls-sapbench-action

YOLO11m-cls — SAP-Bench next-action prediction

Supervised YOLO11m-cls baseline predicting the surgeon's next action on cholecystectomy frames from SAP-Bench.

Trained as a baseline for the SDSC × Chicago Booth surgical video understanding leaderboard (Recommendations tab).

Model

• Ultralytics yolo11m-cls.pt fine-tuned for 5-way single-label classification (Dissection, Vessel Clipping, Tissue Retraction, Coagulation, Aspiration)
• 224×224 inputs, batch 32, up to 100 epochs with patience 15, seed 42, standard color/geometric augmentation
• Full training code in s69_sapbench_supervised.py; training curves in loss_curve.csv

Evaluation

Full 353-frame validation split (95% bootstrap CI):

Metric	Value
Exact-match accuracy	45.6% (40.5–51.3)

See the leaderboard for comparison with vision-language models.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
result = model("frame.jpg")[0]
print(result.names[result.probs.top1])

License note

Derived from Ultralytics YOLO11 weights; this checkpoint is therefore distributed under AGPL-3.0.

Limitations

Research baseline only. Not a medical device. Next-action prediction from a single frame is inherently ambiguous.