AI Learns to Spot Eye Tumors with New Image Dataset

The Challenge of Detecting Eye Tumors Early

Intraocular tumors, which grow inside the eye, can be either benign (non-cancerous) or malignant (cancerous). Benign tumors include conditions like choroidal hemangioma and retinal capillary hemangioma, while malignant ones include dangerous cancers like retinoblastoma and uveal melanoma. While any tumor can impair vision, malignant ones are particularly life-threatening due to their ability to spread to other parts of the body. Unfortunately, these tumors often develop silently, and many patients only seek medical help after their vision has already been significantly affected. This makes early and accurate diagnosis absolutely critical for preserving sight and saving lives.

Fundus photography, a non-invasive method of taking pictures of the back of the eye, is a cornerstone for diagnosing and monitoring eye diseases. Traditional cameras capture a narrow field of view, but modern ultra-wide-field (UWF) imaging systems can capture a panoramic 200° image of the retina. This wide view is a game-changer for detecting intraocular tumors, which often appear in the far periphery of the eye, helping doctors spot lesions earlier and track their growth more effectively.

The AI Training Gap in Ocular Oncology

Artificial intelligence has already shown incredible promise in automatically diagnosing other eye conditions like diabetic retinopathy and glaucoma from fundus images. AI algorithms can analyze thousands of images, learn to spot signs of disease with high accuracy, and help ease the screening burden on doctors. However, progress in using AI for intraocular tumors has been slow. The main reason is a major data bottleneck: a scarcity of large, diverse, and publicly available datasets of tumor images for training these AI models. Most existing datasets are private, focus on a single type of tumor, or use older, narrow-field images, which limits the development of robust and versatile AI tools.

To break this logjam, researchers have developed and released a comprehensive new dataset of UWF fundus images. It covers multiple types of intraocular tumors as well as normal eyes, providing a vital resource to the global research community. This dataset is designed to fuel the development and validation of next-generation AI algorithms, paving the way for more accurate automated diagnostics and better patient outcomes.

Building a High-Quality Medical Dataset

This new dataset contains 2,031 high-quality UWF fundus images collected from Shenzhen Eye Hospital between 2019 and 2024. After removing any blurry or obscured images, the final collection includes 677 images from patients with intraocular tumors and 1,354 from healthy individuals.

To ensure accuracy, every image was meticulously classified by a team of three expert annotators. The images are divided into six distinct categories: normal, Choroidal Hemangioma (CH), Retinal Capillary Hemangioma (RCH), Choroidal Osteoma (CO), Retinoblastoma (RB), and Uveal Melanoma (UM). This diversity is key to training AI models that can differentiate between various conditions.

For researchers and developers eager to use this resource, the complete dataset, titled “UWF Fundus Images of Intraocular Tumors,” is publicly available on Figshare. It is organized into six subfolders, one for each category, making it easy to access and use for building and testing diagnostic algorithms.

Putting the Dataset to the Test with AI Models

To prove the dataset's value, the research team used it to train four different well-known deep learning models: ResNet50, ResNet101, ConvNeXt-T, and ViT-B. The data was split into training, validation, and testing sets, ensuring the models were evaluated on images they had never seen before.

The results were impressive. While all models performed well, the Vision Transformer (ViT-B) model stood out, achieving the highest overall performance with a mean accuracy of 91.46%, an AUC (a measure of diagnostic ability) of 96.87%, and an F1 score of 81.42%. Statistical tests confirmed that the ViT-B model's performance was significantly better than the other three, suggesting that its architecture is particularly well-suited for identifying the subtle features of different intraocular tumors in UWF images.

Real-World Performance and Future Impact

To ensure the AI models weren't just good at identifying tumors from one hospital, the team conducted an external validation using a separate dataset from the EENT Hospital of Fudan University. The models maintained their excellent performance on this new data, demonstrating their robustness and generalizability for real-world clinical use. This strong performance on an external dataset confirms the high quality and utility of the new public dataset.

By making this high-quality, diverse, and expertly-labeled dataset publicly available, the researchers have provided a powerful new tool for the global AI community. It promises to accelerate the development of sophisticated AI systems that can help ophthalmologists detect deadly eye tumors earlier, ultimately leading to better patient care and saving both vision and lives.