license: apache-2.0
tags:
- medical-image-analysis
- oral-leukoplakia
- instance-segmentation
- deep-learning
- pytorch
OLDA-Bench
An Interpretable AI System for Oral Leukoplakia Progression: From Early Screening to Lesion Delineation
Overview
Oral leukoplakia is one of the most common oral potentially malignant disorders (OPMDs) and represents a critical precursor to oral cancer. However, current diagnostic workflows largely rely on invasive biopsy procedures and subjective clinical assessment, limiting their suitability for large-scale screening and longitudinal monitoring.
To address these challenges, we introduce OLDA-Bench (Oral Leukoplakia Diagnosis and Analysis Benchmark), the first benchmark specifically designed for oral leukoplakia progression recognition and interpretable lesion assessment
Dataset Access
The dataset is available through the Hugging Face Dataset Hub:
https://huggingface.co/datasets/gy-chen/OLDA-Bench-Dataset
Access requests are reviewed manually and granted for academic and non-commercial research purposes only.
📦 Dataset Structure
Images identified as Leukoplakia or Leukoplakia with Cancer in the first stage are passed to the segmentation network. This stage uses a dataset in the standard COCO format, structured as follows:
./root_data/
│
├── train/
│ ├── xxx.jpg
│ ├── ...
├── val/
│ ├── xxx.jpeg
│ ├── ...
├── annotations
│ ├── train.json
│ ├── val.json
The instance segmentation task includes two categories:
- Leukoplakia
- Leukoplakia Cancer
The .json annotation files provide pixel-level segmentation masks and category labels for each lesion instance.
📊 Dataset Split
- Training set: 389 images
- Validation set: 44 images
🧪 Benchmark with MMdetection3.x & Detection2
🔹 Training
# Single-GPU training
python tools/train.py\
work_dirs_mask2former_swin_b/mask2former_swin-s-p4-w7-224_8xb2-lsj-50e_coco.py
# Multi-GPU training
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 ./tools/train.py\
work_dirs_mask2former_swin_b/mask2former_swin-s-p4-w7-224_8xb2-lsj-50e_coco.py\
--launcher pytorch
You can replace config with any supported architecture name from mmdetection3.x.
🔹 Evaluation
After training, evaluate a model checkpoint on the test set:
# Single-GPU evaluation
python tools/test.py\
work_dirs_sparseinsts/sparseinst_r50_iam_8xb8-ms-270k_coco.py\
work_dirs_sparseinsts/best_coco_segm_mAP_iter_8500.pth
# Multi-GPU evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 tools/test.py \
work_dirs_sparseinsts/sparseinst_r50_iam_8xb8-ms-270k_coco.py \
work_dirs_sparseinsts/best_coco_segm_mAP_iter_8500.pth \
--launcher pytorch
🔹Additional Info
The OLDA-Bench is based on our previous work (environment code base) published in IEEE JBHI CDTM and OLPR.