How rgonzalezlopez/lisa2025teamupf works
This repository is a MICCAI LISA 2025 solution with two user-facing outcomes: MRI quality/artifact assessment and anatomical segmentation. The README describes ordinal, Bayesian, and ensemble QA, plus nnU-Net and LoRA-fine-tuned SAM-Med3D segmentation. The tree supports these major branches, but sampled implementation excerpts are limited; some internal wiring is therefore shown only where the README or excerpts support it.
Main entry points: src QA workflows · run_model.py Segmentation workflows · sam3D.py 3D prompt segmentation · train_ordinal.py Ordinal QA training
Components 18
Quality assessment
- QA workflows
Task01/src - Ordinal QA training
Task01/src/models/train_ordinal.py - Bayesian QA
Task01/SupCon/validation_bayesian.py - QA ensemble
Task01/SupCon/ensemble_eval.py - Image quality features
Task01/src/classical/feat_extraction.py - Quality and artifacts
Task01/src/evaluation/metrics.py
Segmentation
- Segmentation workflows
run_model.py
nnU-Net runtime
- Dataset planning
nnUNet/nnunetv2/experiment_planning - nnU-Net training
nnUNet/nnunetv2/training/nnUNetTrainer/nnUNetTrainer.py - Sliding-window inference
nnUNet/nnunetv2/inference/predict_from_raw_data.py - Segmentation export
nnUNet/nnunetv2/inference/export_prediction.py - Plans and labels
nnUNet/nnunetv2/utilities/plans_handling/plans_handler.py
SAM-Med3D
- LoRA fine-tuning
SAM-Med3D-lora/train_lora.py - 3D prompt segmentation
SAM-Med3D-lora/segment_anything/modeling/sam3D.py - Interactive clicks
SAM-Med3D-lora/utils/click_method.py - Structure masks
SAM-Med3D-lora/utils/infer_utils.py
Other components
- MRI user
- MRI volumes
Connections 19
- MRI user → MRI volumes: provides
- MRI volumes → QA workflows: assesses
- QA workflows → Ordinal QA training: trains
- QA workflows → Image quality features: extracts features
- Ordinal QA training → Bayesian QA: models co-occurrence
- Bayesian QA → QA ensemble: combines predictions
- QA workflows → Quality and artifacts: reports
- MRI user → Segmentation workflows: requests masks
- MRI volumes → Segmentation workflows: supplies volumes
- Segmentation workflows → Dataset planning: plans data
- Dataset planning → Plans and labels: defines configuration
- Plans and labels → nnU-Net training: configures
- nnU-Net training → Sliding-window inference: provides model
- Sliding-window inference → Segmentation export: exports predictions
- Segmentation workflows → LoRA fine-tuning: fine-tunes alternative
- LoRA fine-tuning → 3D prompt segmentation: adapts model
- 3D prompt segmentation → Interactive clicks: uses prompts
- 3D prompt segmentation → Structure masks: produces masks
- Segmentation export → Structure masks: produces masks