L-MAE: Longitudinal masked auto-encoder with time and severity-aware encoding for diabetic retinopathy progression...

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Self-Supervised Learning

Background:

• Self-supervised learning (SSL) shows promise in computer vision but faces challenges in medical imaging due to unique data characteristics.
• Existing SSL pretext tasks often lack crucial clinical contextual knowledge for decision support.

Purpose of the Study:

• To develop a longitudinal masked auto-encoder (MAE) tailored for medical imaging.
• To enhance MAE with time-aware position embedding and disease progression-aware masking.
• To improve the prediction of disease progression using longitudinal medical data.

Main Methods:

• Developed a Transformer-based longitudinal MAE incorporating time intervals between examinations.
• Introduced a masking strategy that adapts to disease progression during follow-up exams.
• Evaluated the model on the OPHDIAT dataset for diabetic retinopathy (DR) severity prediction.

Main Results:

• The proposed time-aware and disease progression-aware strategies significantly improved predictive accuracy.
• The longitudinal MAE outperformed conventional baseline models and standard longitudinal Transformers.
• The adaptations demonstrated substantial improvements in deep classification model performance for medical domains.

Conclusions:

• Time-aware position embedding and disease progression-informed masking are effective adaptations for longitudinal MAE in medical imaging.
• These methods enhance the capture of temporal trends and pathological changes.
• The developed approach offers a more powerful tool for clinical decision support and disease progression assessment.