Δ t -Mamba3D: A Time-Aware Spatio-Temporal State-Space Model for Breast Cancer Risk Prediction

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning

Background:

• Longitudinal analysis of medical images is challenging due to irregular time intervals.
• Existing models struggle to balance spatial detail and temporal dynamics efficiently.
• Current methods often compromise by losing spatial information or using computationally expensive spatio-temporal models.

Purpose of the Study:

• To develop a novel deep learning architecture for effective longitudinal medical image analysis.
• To address the challenge of modeling high-resolution image sequences captured at non-uniform time steps.
• To improve computational efficiency and accuracy in analyzing sequential radiological data.

Main Methods:

• Introduced Time-Aware Δt-Mamba3D, a state-space architecture for longitudinal medical imaging.
• Incorporated a continuous-time selective scanning mechanism to integrate true time differences.
• Utilized a multi-scale 3D neighborhood fusion module for robust spatio-temporal relationship capture.

Main Results:

• Achieved superior performance in breast cancer risk prediction using sequential mammograms.
• Improved validation C-index by 2-5 percentage points.
• Demonstrated higher 1-5 year AUC scores compared to existing recurrent, transformer, and state-space models.

Conclusions:

• Time-Aware Δt-Mamba3D effectively models irregular time intervals and spatio-temporal context in medical images.
• The model offers a computationally efficient framework for analyzing long patient screening histories.
• This represents a new framework for advanced longitudinal image analysis, particularly in mammography.