Development and validation of an artificial intelligence-based pipeline for predicting oral epithelial dysplasia malignant transformation

Introduction

Oral epithelial dysplasia (OED) presents a significant challenge in the realm of oral pathology, where accurate diagnosis and early detection are paramount for effective intervention and prevention of malignant progression. OED is a potentially malignant histopathological diagnosis encompassing various lesions of the oral mucosa, typically manifesting as white (leukoplakia), red (erythroplakia) or mixed red-white (erythroleukoplakia) lesions^1,2.

Histopathological grading of Haematoxylin and Eosin (H&E) stained tissue using the World Health Organisation (WHO, 2017³) classification system remains the current accepted practice for diagnosis and risk stratification of OED lesions. This is a three-tier system for grading OED into mild, moderate and severe grades based on the presence, severity and location of a wide range of cytological and architectural histological features (28 in total^4,5). By its nature, this approach suffers from significant intra- and inter-observer variability and has poor predictive value for malignant transformation risk, potentially impacting on patient management. An alternate binary grading system, categorising lesions as low- or high-risk, based on the number of cytological and architectural features (as listed in the WHO criteria) aimed to improve the reproducibility of grading^6,7. However, studies have shown significant variability and unreliability in grading using both systems, highlighting the need for more objective and reproducible methods that can better predict malignant transformation risk in OED^8,9.

To address challenges in subjectivity and misclassification of precancerous and cancerous lesions, there is a growing interest in leveraging advanced technologies, particularly deep learning (DL), which has seen extensive use in medical image analysis over the past decade^{10, 11–12}. Several state-of-the-art models, such as U-Net¹³ and DeepLab¹⁴, have been developed to perform image classification and segmentation. These models typically use convolutional neural networks (CNN), such as ResNet¹⁵, as feature extractors. Within digital pathology, weakly supervised methods have became popular choices for the analysis of histology images, enabling slide-level classification based on patch-level predictions. These methods typically divide WSIs into smaller patches, before using CNNs to extract patch-level features^{16, 17–18}. However, despite their success, CNN-based models have limitations such as high computational overhead and difficulty in capturing long-range dependencies in images, when being used for either segmentation or classification.

Transformers have gained widespread attention in...