Advancing oral leukoplakia progression recognition: A benchmark with dataset, method, and application.

Abstract

Oral leukoplakia, a potentially malignant disorder, is a critical precursor to oral squamous cell carcinoma (OSCC), which accounts for 90 % of oral cancer cases. Early recognition of leukoplakia progression is essential for timely intervention and improved patient outcomes. However, most existing methods focus on easy oral cancer recognition, neglecting the nuanced challenge of leukoplakia progression recognition. In this paper, we introduce Oral Leukoplakia Progression Recognition (OLPR) as a novel benchmark task to classify oral lesions into three clinically relevant categories: normal, leukoplakia, and leukoplakia with cancer. We construct the OLPR dataset, a high-quality, annotated collection derived from multiple public datasets, and establish an external validation dataset using university and clinical data. In addition, we propose the Oral Leukoplakia Progression Network (OLPNet), which combines a ConvNeXt backbone pre-trained on large-scale datasets with a Feature Refinement Module (FRM) to enhance feature extraction and refinement for subtle and complex variations in oral lesion progression. Extensive experiments demonstrate the superiority of OLPNet over state-of-the-art methods, achieving 91.34 % and 90.63 % F1-scores on OLPR and external datasets, respectively. Furthermore, we provide a comprehensive benchmark by evaluating over 15 classic and state-of-the-art classification models, offering valuable insights into advancing oral leukoplakia progression recognition. The dataset and related resources are released at https://github.com/qkee-lz/OLPR.