Dental CNN Segmentation

Panoramic dental radiograph detector/segmenter built on YOLOv8 with optional tooth-status classification and FDI numbering. The repo ships only code, the latest detector artifacts, and test history—no bulky caches.

Highlights

• Single entrypoints for detection, FDI numbering, and optional classification (src/pipeline/inference.py).
• Reproducible data prep (VIA → YOLO, merge with COCO) plus benchmarking/plotting helpers in scripts/.
• Latest detector artifacts in workspace/detectors/pano-yolo-merged-cpu/ (PT/ONNX/ONNX-slim + plots).
• Validation snapshots in workspace/validation/pano-yolo-merged-cpu-test6/ with JSON outputs and visual batches.
• Clean Python package under src/ with configs and modular data/model/training components.

Repo Layout

• src/config.py – shared configs + YAML override loader.
• src/data/ – VIA parser, datasets, and Albumentations transforms.
• src/models/ – YOLO wrapper plus ResNet/EfficientNet classifier heads.
• src/training/train_detector.py – YOLOv8 detector training entrypoint.
• src/training/train_classifier.py – optional tooth-status classifier trainer.
• src/pipeline/inference.py – detection + optional classification CLI with FDI numbering.
• scripts/prepare_yolo_dataset.py – convert VIA annotations into YOLO folders and dental.yaml.
• scripts/build_merged_yolo_dataset.py – merge VIA and Roboflow COCO datasets.
• scripts/benchmark_detector.py, scripts/plot_training_metrics.py – runtime and metric plotting helpers.
• workspace/ – latest detector + validation history kept after cleanup.

Plots from the latest run:

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
export PYTHONPATH=$(pwd)

Data Preparation

1. Convert VIA labels to YOLO segmentation:

PYTHONPATH=$(pwd) python scripts/prepare_yolo_dataset.py \
  --images-dir data/via \
  --annotations data/via/annotations.json \
  --output-dir workspace/yolo \
  --task segment --train-ratio 0.75 --val-ratio 0.15 --seed 1337

2. (Optional) Merge with the Roboflow COCO export stored in data/coco/:

PYTHONPATH=$(pwd) python scripts/build_merged_yolo_dataset.py \
  --coco-root data/coco \
  --via-images data/via \
  --via-annotations data/via/annotations.json \
  --output-dir workspace/yolo_merged

Training and Evaluation

• Train detector (YOLOv8s-seg, 80 epochs, 1024px, etc.):

PYTHONPATH=$(pwd) python -m src.training.train_detector \
  --data-yaml workspace/yolo/dental.yaml \
  --model yolov8s-seg.pt \
  --project workspace/detectors \
  --name pano-yolo-merged-cpu

• Plot metrics for any run:

PYTHONPATH=$(pwd) python scripts/plot_training_metrics.py \
  --results-csv workspace/detectors/pano-yolo-merged-cpu/results.csv \
  --output-dir workspace/detectors/pano-yolo-merged-cpu

• Benchmark ONNX/pt weights:

PYTHONPATH=$(pwd) python scripts/benchmark_detector.py \
  --weights workspace/detectors/pano-yolo-merged-cpu/weights/best.onnx \
  --images workspace/validation/pano-yolo-merged-cpu-test6/val_batch1_pred.jpg

Inference

Run detection only (FDI numbering included):

PYTHONPATH=$(pwd) python -m src.pipeline.inference \
  --images "data/via/*.jpg" \
  --detector workspace/detectors/pano-yolo-merged-cpu/weights/best.onnx \
  --skip-classifier \
  --output-json workspace/validation/detections_only.json

Add --classifier path/to/classifier.pt (drop the skip flag) to enable status prediction and probability dumps.

Current Results

• Best box epoch (71): box mAP50 0.959, mAP50-95 0.674; mask mAP50 0.947, mAP50-95 0.573 (workspace/detectors/pano-yolo-merged-cpu/results.csv).
• Final epoch (80) masks: precision 0.943, recall 0.913, mAP50 0.938.
• Visual artifacts live in workspace/detectors/pano-yolo-merged-cpu/*.png and workspace/validation/pano-yolo-merged-cpu-test6/*.jpg.

Sample predictions:

Housekeeping

• Temporary caches and older runs were removed; regenerate splits via the prep scripts if needed.
• Keep workspace/detectors/pano-yolo-merged-cpu/ and workspace/validation/ to preserve the referenced history.
• workspace/validation/detections_only.json still references former dataset/ paths; adjust if replaying it.