SSAD: Self-supervised Auxiliary Detection Framework for Panoramic X-ray based Dental Disease Diagnosis
This is the offical implement code of artical https://arxiv.org/abs/2406.13963. It employs the reconstruction-based Self-supervised methods to assist deep-learning model in diagnosing the Dental Disease, and the whole framework as below:
# STEP1: Find a PyTorch version that matches your CUDA version from the official PyTorch website, for example:
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
# STEP2: Install Requirements
pip install -r requirements.txt
# STEP3 Install CLIP
cd models/CLIP && pip install .
# STEP4 Refer to DINO for guidance on installing MultiScaleDeformableAttention.
DINO website is there
Data Preparation: Please refer at there
## Train SSAD with detectron2, For Example:
# yaml
python train_diffdet.py --num-gpus 1 --config-file configs/faster_rcnn/diffdet.dentex.swinbase.disease.yaml OUTPUT_DIR "the path of model weights" MODEL.WEIGHTS "checkpoints/swin_base_patch4_window7_224_22k.pth"
# py
python train_lazyconfig_net.py --num-gpus 1 --config-file configs/faster_rcnn/faster_rcnn_vitB_vitdet.py "train.output_dir='the path of model weights'" "train.init_checkpoint='checkpoints/vitdet.pkl'"
## Train Yolov8, For Example:
cd yolov8-ssad
# Run Baseline
export DEFAULT_CFG_PATH=./hyp.baseline.yaml && python train.py
# Run SSAD
export DEFAULT_CFG_PATH=./hyp.ssad.yaml && python train_ssad.py
| Network | Encoder | SSAD | checkpoint | AP50:95 |
|---|---|---|---|---|
| DINO | Res50 | N | 15.42 | |
| DINO | Res50 | Y | 17.12 | |
| YOLOv8-L | CSPDarknet | N | 33.9 | |
| YOLOv8-L | CSPDarknet | Y | 37.0 | |
| Faster RCNN | ViTDet-B | N | 28.93 | |
| Faster RCNN | ViTDet-B | Y | 31.04 | |
| FCOS | ViTDet-B | N | 14.46 | |
| FCOS | ViTDet-B | Y | 19.18 | |
| DiffusionDet | Swin-B | N | 10.25 | |
| DiffusionDet | Swin-B | Y | 11.40 | |
| HierarchicalDet | Swin-B | N | 24.15 | |
| HierarchicalDet | Swin-B | Y | 26.70 |