MARTA CX43 slide classifier
This repository hosts the pretrained checkpoint of the MARTA classifier used to detect ROI patterns associated with longitudinal versus lateralized CX43 distributions in cardiac tissue images.
Files
best_stage3_full.pth
Final trained model checkpoint.
Model summary
This model corresponds to the slide-level training setup used in the main MARTA experiments.
Architecture
- input mode:
stack - fusion:
dual - head:
mlp - hidden size:
256 - dropout:
0.5
Training setup
- resize:
384 - stage 1 epochs:
6 - stage 2 epochs:
4 - stage 3 epochs:
10 - partially unfrozen backbone blocks in stage 2:
2
Optimization
- head learning rate:
9.83e-4 - last-block learning rate:
3.44e-4 - rest-of-backbone learning rate:
1.15e-4 - weight decay:
1e-4
Data split
The model was trained and evaluated using the precomputed slide-level split:
- train slides:
IM6,IM9,IM133 - validation slide:
IM1315 - test slides:
IM1313,IM1314
This split was chosen to reduce leakage across correlated ROIs from related slides.
Performance summary
Performance is reported for the selected best-validation checkpoint of the final training run.
Validation
- validation loss:
0.0599 - ROC-AUC:
0.8022 - PR-AUC:
0.5924 - precision (class 1):
0.5000 - recall (class 1):
0.5000 - precision (class 0):
0.8200 - recall (class 0):
0.8200
Test
- ROC-AUC:
0.9048 - PR-AUC:
0.8102 - precision (class 1):
0.7273 - recall (class 1):
0.6667 - precision (class 0):
0.9036 - recall (class 0):
0.9259
Split sizes
- train:
2966 - validation:
68 - test:
105
Usage
This checkpoint is intended to be used with the MARTA repository.
After downloading the file, place it at:
trained_model/best_stage3_full.pth
Then run inference from the main MARTA repository using the standard inference commands described in its README.
##Notes This repository only hosts the pretrained checkpoint. Code, documentation, inference pipeline, and reproducibility artifacts are available in the main MARTA GitHub repository.