How mathpluscode/cinema works
CineMA is a cardiac MRI foundation-model package: dataset-specific preprocessing prepares imaging data, masked-autoencoder pretraining produces a backbone, and task-specific training and evaluation support segmentation, classification, EF/other regression, and landmark localization. The graph includes the shared model/training layers and the documented dataset and task branches. Several dataset and task modules are unsampled, so their links are limited to capabilities explicitly documented in the README or references recorded in the source index.
Main entry points: segmentation/train.py Segmentation training · cinema/train.py Training runtime · classification Classification tasks · data Dataset preprocessing
Components 22
Imaging data
- Dataset preprocessing
cinema/data - DICOM handling
cinema/data/dicom.py - Image transforms
cinema/data/sitk.py
Foundation model
- MAE pretraining
cinema/mae/pretrain.py - CineMA encoder-decoder
cinema/mae/mae.py - Vision Transformer
cinema/vit.py - Convolution blocks
cinema/conv.py - ConvViT
cinema/convvit.py
Downstream tasks
- Segmentation training
cinema/segmentation/train.py - ConvUNetR
cinema/segmentation/convunetr.py - Segmentation dataset
cinema/segmentation/dataset.py - Segmentation evaluation
cinema/segmentation/eval.py - Classification tasks
cinema/classification - Classification dataset
cinema/classification/dataset.py - Regression tasks
cinema/regression - Landmark localization
cinema/segmentation/landmark
Shared runtime
- Clinical metrics
cinema/metric.py - Training transforms
cinema/transform.py - Training runtime
cinema/train.py - Device selection
cinema/device.py - Logging
cinema/log.py
Other components
- Researcher
Connections 25
- Researcher → Dataset preprocessing: prepares datasets
README.md - Dataset preprocessing → Image transforms: uses image operations
cinema/data/mnms/preprocess.py - Dataset preprocessing → DICOM handling: uses DICOM utilities
cinema/data/rescan/preprocess.py - Researcher → MAE pretraining: pretrains backbone
README.md - MAE pretraining → CineMA encoder-decoder: trains
cinema/mae/pretrain.py - CineMA encoder-decoder → Vision Transformer: uses transformer
cinema/mae/mae.py - CineMA encoder-decoder → Convolution blocks: uses convolutions
cinema/mae/mae.py - Researcher → Segmentation training: fine-tunes
README.md - Segmentation training → ConvUNetR: uses model
cinema/segmentation/train.py - Segmentation training → Training transforms: uses transforms
cinema/segmentation/train.py - Segmentation training → Training runtime: uses runtime
cinema/segmentation/train.py - Segmentation dataset → Image transforms: reads image data
- Segmentation evaluation → Segmentation dataset: loads samples
cinema/segmentation/eval.py - Segmentation evaluation → Clinical metrics: computes metrics
cinema/segmentation/eval.py - Researcher → Classification tasks: trains or evaluates
README.md - Classification tasks → Classification dataset: loads samples
cinema/classification/acdc/train.py - Classification tasks → Training runtime: uses runtime
cinema/classification/acdc/train.py - Classification tasks → ConvViT: builds model
cinema/classification/train.py - Researcher → Regression tasks: trains or evaluates
README.md - Regression tasks → Training runtime: uses runtime
cinema/regression/acdc/train.py - Regression tasks → Clinical metrics: uses metrics
cinema/regression/landmark/train.py - Researcher → Landmark localization: localizes landmarks
README.md - Landmark localization → Segmentation training: uses segmentation framework
cinema/segmentation/landmark/train.py - Training runtime → Device selection: selects device
cinema/train.py - Training runtime → Logging: initializes logging
cinema/train.py