Project case study · Medical AI · 3D MRI · Deep learning
3D cardiac MRI segmentation with MONAI and a 3D U-Net
This project segments cardiac structures in volumetric MRI. It documents a complete research pipeline from NIfTI loading and preprocessing through mixed-precision training, sliding-window inference, and validation-set Dice evaluation.
01 · Context
Research question
Cardiac MRI volumes contain three-dimensional spatial context, so the workflow needs to preserve volumetric information through preprocessing, training, and inference. The project focuses on a public benchmark task and transparent reporting of its validation result.
02 · Methods
Approach
- Used the Medical Segmentation Decathlon Task02 Heart dataset and NIfTI volumes.
- Built preprocessing and data-loading steps with PyTorch and MONAI.
- Trained a 3D U-Net with DiceCELoss and AdamW for 100 epochs, using mixed precision (FP16).
- Applied sliding-window inference to full 3D volumes and evaluated voxel-level overlap with Dice.
03 · Evaluation
Reported results
- Best reported validation Dice: 0.8821 on the documented benchmark split.
- Training configuration: 100 epochs, 3D U-Net, mixed-precision training.
- The result is a project-reported benchmark validation metric, not a clinical performance estimate.
04 · Scope
Limitations
This work has not been clinically validated, prospectively tested, or evaluated across external institutions and scanner settings. It is not a diagnostic device or clinical decision-support system. The validation score describes only the project’s benchmark split.