M2Net: Multimodal Multitask Mutual Learning for Anti-VEGF Efficacy Prediction.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Neovascular age-related macular degeneration (AMD) is a primary cause of vision loss in older adults.
• Current anti-VEGF treatments for AMD are costly and their effectiveness varies.
• Accurate prediction of treatment efficacy is vital for personalized patient management.

Purpose of the Study:

• To develop a novel AI framework, M2Net, for simultaneously predicting visual acuity changes and generating post-treatment retinal images.
• To improve the prediction of anti-VEGF treatment efficacy in neovascular AMD.
• To integrate multimodal imaging data for more robust predictions.

Main Methods:

• M2Net utilizes a dual-branch network processing fundus photographs and Optical Coherence Tomography (OCT) scans.
• A Multimodal Multitask Mutual learning approach is employed.
• Key innovations include modules for collaborative multimodal feature interaction and joint pre-post treatment image analysis.

Main Results:

• M2Net achieved a classification accuracy of 96.03% in predicting treatment efficacy.
• The model demonstrated strong performance in image generation, with SSIM scores of 0.6377 for OCT and 0.8347 for fundus images.
• Experimental results on the multimodal MMPD dataset show superior performance over existing methods.

Conclusions:

• M2Net offers a significant advancement in predicting neovascular AMD treatment outcomes by jointly analyzing multimodal imaging data.
• The developed framework enhances prediction accuracy and image generation capabilities.
• This approach holds promise for improving patient care and managing neovascular AMD more effectively.