DDaTR: Dynamic Difference-Aware Temporal Residual Network for Longitudinal Radiology Report Generation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Radiology Report Generation (RRG) automates report creation from medical images.
• Longitudinal Radiology Report Generation (LRRG) enhances RRG by comparing current and prior exams to track temporal changes.
• Existing LRRG methods struggle to capture spatial and temporal correlations, leading to suboptimal performance.

Purpose of the Study:

• To develop a novel network, the dynamic difference-aware temporal residual network (DDaTR), for improved LRRG.
• To effectively capture multi-level spatial correlations and temporal dynamics in longitudinal medical imaging data.

Main Methods:

• Introduced two modules within the visual encoder: Dynamic Feature Alignment Module (DFAM) for prior feature integrity and dynamic difference-aware module (DDAM) for capturing inter-exam differences.
• Employed a dynamic residual network for unidirectional transmission of longitudinal information to model temporal correlations.
• Evaluated DDaTR on three benchmarks for both RRG and LRRG tasks.

Main Results:

• DDaTR demonstrated superior performance compared to existing methods on both RRG and LRRG tasks.
• The proposed modules effectively captured spatial correlations and temporal dynamics, improving the representation of changes across exams.
• The network successfully modeled longitudinal information, leading to more accurate radiology report generation.

Conclusions:

• The DDaTR network offers a significant advancement in Longitudinal Radiology Report Generation.
• The approach effectively addresses limitations in capturing temporal correlations and differences between medical imaging exams.
• DDaTR shows strong efficacy for both RRG and LRRG, paving the way for more accurate automated radiology reporting.