zagari/sentinel-surgical-yolov8m-bleeding

Sentinel Surgical — YOLOv8m for Bleeding Detection in Gynecological Surgery

YOLOv8m fine-tuned to detect anomalous bleeding and surgical instruments in laparoscopic gynecological surgeries. Part of Sentinel Health, a FIAP Tech Challenge Phase 4 academic project on multimodal monitoring of women's health.

This is the v3_finetuned checkpoint — the production model selected after a 3-phase cross-dataset training pipeline.

Performance

Cross-dataset validation on the GynSurg Action Recognition dataset using a fixed validation set of 20 clips (10 bleeding + 10 non-bleeding, ~1800 frames total).

Metric	Value
Bleeding detection rate	91.72%
False positive rate	13.44%
Confidence threshold (recommended)	0.30
Input size	640 × 640

Threshold sweep on the validation set

Threshold	Detection rate	False positive rate	Use case
0.10	94.48%	19.33%	Maximum sensitivity
0.20	93.05%	14.78%	High sensitivity
0.30	91.72%	13.44%	Balanced (default)
0.40	90.07%	10.67%	Conservative balance
0.50	88.96%	9.33%	Lower FP
0.60	87.86%	6.78%	Best (det − FP) score
0.70	85.76%	4.89%	Minimum FP

For surgical contexts, the project chose threshold 0.30 — preferring to over-detect rather than miss real bleeding events.

Classes

ID	Class	Description
0	grasper	Surgical grasping forceps
1	blood	Anomalous bleeding region

Training pipeline (3 phases)

Phase 1 — Baseline

• Dataset: CholecSeg8k — 8,080 frames of laparoscopic cholecystectomy with pixel-level segmentation masks
• Split: 6,464 train / 1,616 val
• Hyperparams: YOLOv8m base, 100 epochs, batch 16, img 640, patience 20
• Result: 5.41% detection · 76.11% FP rate

Phase 2 — Class weighting

• Same data + same hyperparams but with cls=3.0 to compensate the severe class imbalance (5.4 : 1 grasper-vs-blood instances)
• Result: 12.14% detection · 46.89% FP rate

Phase 3 — Cross-dataset fine-tuning (this checkpoint)

• Base: v2 model from Phase 2
• Dataset: 1,020 manually annotated frames from GynSurg (475 bleeding + 545 non-bleeding) using a dedicated web annotation interface built into Sentinel Health
• Technique: Pseudo bounding boxes (central 80% of frame for bleeding labels)
• Hyperparams: 30 epochs, batch 8, lr0 0.001, freeze 10 layers
• Result: 91.72% detection · 13.44% FP rate ✅ — meets both project targets (>60% detection, <20% FP)

Datasets

CholecSeg8k (training)

• Type: Cholecystectomy laparoscopic videos with pixel-level segmentation
• Volume: 8,080 frames
• Instances: 13,680 grasper · 2,545 blood
• Source: Kaggle — CholecSeg8k
• Purpose here: training the underlying object-detection capability

GynSurg Action Recognition (validation + fine-tuning)

• Type: 3-second clips of laparoscopic gynecological surgery
• Resolution: 3840 × 2160 @ 30 fps
• Subsets used: 977 bleeding clips, 1,064 non-bleeding clips
• Source: Medical University of Vienna / University of Toronto
• License: CC BY-NC-ND 4.0
• Purpose here: cross-dataset validation and fine-tuning to ginecological domain

Usage

Direct via Ultralytics

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model_path = hf_hub_download(
    repo_id="zagari/sentinel-surgical-yolov8m-bleeding",
    filename="best.pt",
)
model = YOLO(model_path)
results = model("frame.jpg", conf=0.30)

Direct download (no Python deps)

curl -fL -o best.pt \
  https://huggingface.co/zagari/sentinel-surgical-yolov8m-bleeding/resolve/main/best.pt

In Sentinel Health platform

The Sentinel Surgical container auto-downloads this model from Hugging Face Hub on first start via its entrypoint script — no manual setup required.

Intended use

• ✅ Academic research and demonstrations on surgical video understanding
• ✅ Baseline for further fine-tuning on ginecological surgery datasets
• ✅ Educational illustrations of cross-dataset transfer learning (general → specific surgical domain)

⚠️ Limitations — academic prototype, NOT a medical device

Sentinel Surgical is an academic prototype developed as part of the Tech Challenge — Phase 4 of the postgraduate program in Artificial Intelligence for Developers at FIAP (Faculdade de Informática e Administração Paulista), Brazil.

It MUST NOT be used for:

• Clinical decisions
• Diagnosis
• Surgical guidance in real procedures
• Any safety-critical application

Known limitations:

• Trained on research datasets with biases inherent to specific surgical centers and patient populations
• Validation set is small (20 clips, ~1,800 frames) — generalization beyond similar gynecological laparoscopy is not characterized
• The "blood" class was learned from cholecystectomy (Phase 1-2) and adapted via pseudo-bbox labels (Phase 3) — bounding box localization should not be interpreted as precise tissue segmentation
• False positives at 13.44% means roughly 1 in 7 non-bleeding clips would still raise an alert — a human-in-the-loop is required

Any interpretation of system outputs requires qualified medical professionals.

Team — Grupo Sala 14

• Adriana Martins de Souza — RM 368050
• Diego Oliveira da Silva — RM 367964
• Eduardo Nicola F. Zagari — RM 368021
• Renan de Assis Torres — RM 368513

License

MIT — see LICENSE in the main repository.

Citation

If you use this model in academic work, please cite:

@misc{sentinel_surgical_2026,
  author = {Souza, Adriana M. de and Silva, Diego O. and Zagari, Eduardo N. F. and Torres, Renan A.},
  title = {Sentinel Surgical — YOLOv8m for Bleeding Detection in Gynecological Surgery},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/zagari/sentinel-surgical-yolov8m-bleeding}},
  note = {FIAP Tech Challenge Phase 4 — multimodal monitoring of women's health}
}