How alena-xinran/mri2rep works
MRI2Rep turns free-text liver MRI reports into structured lesion labels, then trains a 3D CNN–Transformer to generate those structured reports from three-channel MRI volumes. Its user-facing web app accepts MRI inputs and presents generated reports, with validation-set comparison and visualization features. The extraction flow is visible in source; model training, evaluation, pretraining, and web capabilities are also included where documented in the README. Some subsystem links rely on README descriptions or source-index references rather than complete excerpts.
Main entry points: app.py Inference and web routes · main.py Training orchestrator · dataset.py MRI sequence dataset · model.py MRI report generator
Components 25
Report labeling
- Batch runner
Report_Process/main.py - Report and prompt I/O
Report_Process/utils/io.py - Label extraction
Report_Process/pipeline.py - LLM client
Report_Process/utils/llm.py - Case logs
Report_Process/utils/logger.py - Label normalization
Report_Process/post.py - Consensus recovery
Report_Process/recover_2of3.py - Failed-case reruns
Report_Process/rerun_failed.py - Dataset aggregation
Report_Process/build_accurate_dataset.py - Batch summary merge
Report_Process/merge_summary_yes.py
Model and training
- Training orchestrator
main.py - Model configuration
src/config.py - MRI sequence dataset
src/dataset.py - MRI report generator
src/model.py - Training and metrics
src/engine.py - PV pretraining
pretrain_pv.py - Checkpoints and logs
Evaluation and analysis
- NLG evaluation
evaluate_nlg.py - Prediction analysis
analyze_preds.py - GradCAM visualization
gradcam_vis.py
Web application
- Inference and web routes
webapp/app.py
Other components
- Report inputs
- MRI web user
- MRI volumes
- Structured report
Connections 36
- Report inputs → Batch runner: provides reports
- Batch runner → Report and prompt I/O: finds and reads
Report_Process/main.py - Batch runner → Case logs: logs cases
Report_Process/main.py - Batch runner → Label extraction: runs stages
Report_Process/main.py - Label extraction → Report and prompt I/O: loads prompts
Report_Process/pipeline.py - Label extraction → LLM client: requests completions
Report_Process/pipeline.py - LLM client → Case logs: records responses
- Batch runner → Label normalization: formats outputs
- Consensus recovery → Failed-case reruns: selects reruns
- Failed-case reruns → Label extraction: reprocesses cases
Report_Process/rerun_failed.py - Label normalization → Dataset aggregation: prepares labels
- Dataset aggregation → Batch summary merge: produces summaries
- Batch summary merge → MRI sequence dataset: supplies training labels
- Training orchestrator → MRI sequence dataset: loads splits
main.py - Training orchestrator → MRI report generator: optimizes model
main.py - Training orchestrator → Training and metrics: trains and evaluates
main.py - Training orchestrator → Model configuration: uses settings
main.py - Training orchestrator → Checkpoints and logs: saves checkpoints
- PV pretraining → MRI report generator: pretrains backbone
pretrain_pv.py - PV pretraining → Model configuration: uses settings
pretrain_pv.py - NLG evaluation → MRI sequence dataset: loads evaluation data
evaluate_nlg.py - NLG evaluation → MRI report generator: generates reports
evaluate_nlg.py - NLG evaluation → Training and metrics: uses evaluation metrics
evaluate_nlg.py - NLG evaluation → Checkpoints and logs: loads checkpoint
- Prediction analysis → MRI sequence dataset: loads cases
analyze_preds.py - Prediction analysis → MRI report generator: compares predictions
analyze_preds.py - Prediction analysis → Checkpoints and logs: loads run
- GradCAM visualization → MRI sequence dataset: loads MRI cases
gradcam_vis.py - GradCAM visualization → MRI report generator: computes attribution
gradcam_vis.py - GradCAM visualization → Checkpoints and logs: loads checkpoint
- MRI web user → Inference and web routes: submits MRI
webapp/app.py - MRI volumes → Inference and web routes: uploads volumes
webapp/app.py - Inference and web routes → MRI sequence dataset: loads validation cases
webapp/app.py - Inference and web routes → MRI report generator: runs inference
webapp/app.py - Inference and web routes → Checkpoints and logs: loads checkpoint
webapp/app.py - Inference and web routes → Structured report: formats findings
webapp/app.py