Spatiotemporal Deformable Transformer for Lung Cancer Risk Prediction Using Low-Dose CT Images

Spatiotemporal Deformable Transformer for Lung Cancer Risk Prediction Using Low-Dose CT Images

Low-dose CT (LDCT) screening saves lives by catching lung cancer early, but interpreting thousands of slices per scan overwhelms radiologists and produces inconsistent risk assessments. This project delivers a state-of-the-art vision transformer medical imaging system that models the full three-dimensional context of lung scans. The architecture employs a spatiotemporal deformable transformer that adaptively focuses attention on relevant anatomical regions across both spatial dimensions and sequential CT slices — capturing how nodules evolve across the scan volume, a decisive factor in malignancy assessment. Deformable attention mechanisms allow the model to track irregular, moving nodule boundaries without rigid grid constraints, while temporal modeling distinguishes stable benign findings from suspicious growth patterns. The system outputs calibrated per-patient risk scores, nodule localization, and attention visualizations that align with radiological reasoning. Trained and evaluated on standard LDCT screening datasets and implemented with PyTorch and MONAI, this project represents cutting-edge lung cancer AI research — ideal for medical imaging postgraduates and developers of clinical decision support AI.

Components

Python 3.8+
PyTorch
MONAI
Deformable attention modules
LDCT screening datasets
Attention visualization tools

Key Features

• Full 3D spatiotemporal modeling of CT volumes across sequential slices
• Deformable attention tracking irregular, moving nodule boundaries without rigid grids
• Temporal modeling distinguishes stable benign findings from suspicious growth
• Calibrated per-patient risk scores with nodule localization
• Attention visualizations aligned with radiological reasoning

Applications

Medical imaging postgraduates, clinical decision support developers and 3D medical image analysis researchers.