Attention-Based Multimodal Deep Learning for Uveal Melanoma Classification Using Ultra-Widefield Fundus Images and...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Uveal melanoma (UM) and choroidal nevi are common intraocular tumors requiring accurate differentiation.
• Current diagnostic methods may have limitations in distinguishing between these conditions.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated classification of uveal melanoma and choroidal nevi.
• To assess the efficacy of integrating ultra-widefield fundus photography and B-scan ultrasonography.

Main Methods:

• A retrospective study included 174 patients with UM or choroidal nevi.
• Deep learning models were trained using fundus photography, ultrasound images, or a combination of both.
• Fivefold cross-validation was employed to evaluate model performance.

Main Results:

• The model combining fundus photography and ultrasound achieved the highest accuracy (94%), F1 score (0.9445), and AUC (0.9606).
• Single-modality models showed strong performance, with transverse ultrasound yielding the best results among them (accuracy: 92%).
• The multimodal model effectively integrated complementary information, outperforming single-modality approaches.

Conclusions:

• Multimodal deep learning integrating fundus photography and ultrasound imaging significantly improves the classification of UM and choroidal nevi.
• This approach demonstrates the feasibility of leveraging multiple imaging modalities for automated diagnosis.
• The developed model shows promise for enhancing the diagnostic accuracy of ocular tumors.