A domain-specific pretrained model for detecting malignant and premalignant ocular surface tumors: A multicenter model development and evaluation study

Background

Malignant and premalignant ocular surface tumors (OSTs) can be sight- and life-threatening, potentially leading to vision loss, cosmetic disfigurement, severe morbidity, and even death. Early detection and appropriate treatment of malignant and premalignant OSTs can improve patient prognosis, reducing cancer-related mortality and leading to optimal functional outcomes. However, atypical variants of these tumors, combined with a shortage of experienced ophthalmologists, may result in delayed diagnosis and treatment. Although medical artificial intelligence (AI) has shown considerable promise in aiding ophthalmologists with the early detection of various eye diseases, training traditional convolutional neural networks (CNNs) for OST diagnosis presents significant challenges due to the limited availability of large, well-annotated datasets with histopathologically labeled images of OSTs.

Research Progress

To address the above issues, Prof. Zhongwen Li's team at Ningbo Eye Institute, Wenzhou Medical University, introduced the Ocular Surface Pretrained Model (OSPM), a novel self-supervised learning-based domain-specific model to mitigate the scarcity of labeled data. It leverages 0.76 million unlabeled ocular surface images from 10 clinical centers across China to learn ocular surface feature representations, which can be easily adapted for classifying OSTs. Subsequently, they developed and evaluated an OSPM-enhanced classification model (OECM) using 1,455 histopathologically labeled OST images to differentiate between malignant, premalignant, and benign OSTs (Fig. 1). OECM demonstrated robust performance (with all AUCs exceeding 0.891) on images collected from multiple clinical centers using different imaging devices (Fig. 2). OECM also performed effectively on images captured with standard digital cameras. This indicates that OECM could be applied to such cameras, offering a cost-effective and convenient approach for high-risk groups to proactively detect malignant and premalignant OSTs through self-photography. Moreover, they confirmed that OECM exhibited higher label efficiency and training efficiency compared to CNN models. This indicates that OECM has the potential to reduce the annotation workload for experts, perform well in tasks with limited data, and offer excellent fine-tuning efficiency. Finally, they validated that OECM can help junior ophthalmologists improve their diagnostic accuracy in OSTs (Fig. 3), potentially enabling them to deliver prompt and effective treatment strategies for patients with malignant and premalignant tumors.

Future Prospects

The research team is advancing OECM with three primary application scenarios:

▶Tertiary Hospitals: To assist in rapid initial screening, significantly reducing patient wait times.

▶County-Level Hospitals: To enhance diagnostic confidence among clinicians in primary care settings, minimizing unnecessary referrals and missed cases.

▶High-Risk Populations: To deploy a mobile phone-based self-check system enabling "test early, detect early" through photographic screening.

The research team envisions OECM becoming a digital sentinel for ocular surface oncology. By enabling timely identification and triage of precancerous lesions and malignant tumors, the system aims to improve both survival rates and long-term visual function outcomes for patients.

Sources：https://spj.science.org/doi/10.34133/research.0711