SurgBox
Collection
Instrument bounding boxes for Cholec80 and the detectors trained on them (MICAD 2026). • 5 items • Updated
You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This detector was trained on Cholec80, which is distributed by CAMMA under its own terms. It may be used for non-commercial purposes only (CC BY-NC-SA 4.0), and its use is also subject to the Ultralytics license (AGPL-3.0).
Log in or Sign Up to review the conditions and access this model content.
SurgBox-RT-DETR-L
RT-DETR-L detector for the seven Cholec80 surgical instruments, trained only on SurgBox pseudo-labels, with no human box used in training.
Paper: From Presence Labels to Bounding Boxes: Can Multimodal Large Language Models Scale Surgical Instrument Localization? (MICAD 2026)
Code: github.com/M-Hamdy-M/SurgBox
Dataset: M-Hamdy/SurgBox
Model
| Architecture | RT-DETR-L (Ultralytics), COCO-pretrained |
| Classes | Grasper, Bipolar, Hook, Scissors, Clipper, Irrigator, SpecimenBag |
| Training data | SurgBox train split (56 Cholec80 videos) |
| Training | 640 px, 20 epochs, default augmentation |
Results
COCO mAP on the 16 held-out test videos (39,354 frames) and on the expert-verified subset of the human reference (197 frames, 238 boxes), paper Table 2.
| mAP | AP50 | AP75 | F1 | |
|---|---|---|---|---|
| Test set | 0.710 | 0.869 | 0.781 | 89.8 |
| Expert-verified subset | 0.839 | 0.937 | 0.901 | 88.3 |
Usage
from huggingface_hub import hf_hub_download
from ultralytics import RTDETR
model = RTDETR(hf_hub_download("M-Hamdy/SurgBox-RT-DETR-L", "rtdetr-l_surgbox.pt"))
results = model.predict("frame.png", imgsz=640, conf=0.25)
for box in results[0].boxes:
print(model.names[int(box.cls)], box.xyxy[0].tolist(), float(box.conf))
The model was trained on Cholec80 frames (854×480). Cholec80 is not redistributed; request it from CAMMA.
License
Released for non-commercial use only, under CC BY-NC-SA 4.0, consistent with Cholec80. The model was trained with Ultralytics, whose license (AGPL-3.0) also applies to its use.
Citation
If you use this model, please cite our paper:
@inproceedings{hamdy2026surgbox,
title = {From Presence Labels to Bounding Boxes: Can Multimodal Large Language Models Scale Surgical Instrument Localization?},
author = {Hamdy, Mohamed and Abdel-Ghani, Muraam and Ahmed, Fatmaelzahraa and Nasar, Sifna and Ahmed, Mariam and Al-Jalham, Khalid and Al-Ali, Abdulaziz and Balakrishnan, Shidin},
booktitle = {Medical Imaging and Computer-Aided Diagnosis (MICAD)},
year = {2026}
}
Contact
Questions, issues and suggestions are welcome. Please open a discussion on this page or contact me.
- Downloads last month
- 3