akifkaradag/cardiac-segmentation-diagnosis

Cardiac Segmentation & Heart Disease Diagnosis (ACDC 3D nnU-Net)

An end-to-end medical AI pipeline for automated 3D cardiac segmentation, clinical biomarker derivation (Ejection Fraction, Ventricular Volumes, Myocardial Mass), and 5-class cardiac disease diagnosis from short-axis cine cardiac MRI scans.

• GitHub Repository: Makifkaradag/cardiac-segmentation-diagnosis
• Framework: nnU-Net v2 (3d_fullres) & Random Forest Classifier
• Benchmark Dataset: Automated Cardiac Diagnosis Challenge (ACDC, MICCAI)

---

Model Summary

• Task 1 (Segmentation): Full 3D segmentation of three cardiac structures:
  • Label 1: Right Ventricle Cavity (RV)
  • Label 2: Left Ventricle Myocardium (MYO)
  • Label 3: Left Ventricle Cavity (LV)
• Task 2 (Clinical Biomarker Extraction): Automatically derives Left/Right Ventricular End-Diastolic Volume (EDV), End-Systolic Volume (ESV), Ejection Fraction (EF %), and Myocardial Mass (g).
• Task 3 (Diagnosis Classification): 5-class pathology classification:
  • NOR: Normal cardiac function
  • MINF: Previous Myocardial Infarction
  • DCM: Dilated Cardiomyopathy
  • HCM: Hypertrophic Cardiomyopathy
  • RV: Abnormal Right Ventricle

---

Benchmark Performance & Evaluation

Evaluated on 100 ACDC training patients (5-fold cross-validation) and 50 holdout cases.

1. 3D Segmentation Quality (Holdout Set)

Anatomical Structure	Target Region	Mean Dice Score
Left Ventricle (LV)	Blood cavity	0.945
Right Ventricle (RV)	Crescent blood cavity	0.911
Myocardium (MYO)	Left ventricular wall	0.900

2. Clinical Biomarker Agreement (Ground Truth vs. Predictions)

Clinical Metric	Mean Absolute Error (MAE)	Pearson Correlation (r)
LV Ejection Fraction (%)	2.31	0.990
RV Ejection Fraction (%)	5.17	0.889
Myocardial Mass (g)	7.46	0.982

3. Diagnosis Classification (5-Fold Stratified Cross-Validation)

• Overall Accuracy: 0.880 (88.0%)

Disease Class	Precision	Recall	F1-Score	Support
DCM	0.95	0.95	0.95	20
HCM	1.00	0.80	0.89	20
MINF	0.90	0.95	0.93	20
NOR	0.72	0.90	0.80	20
RV	0.89	0.80	0.84	20

• End-to-End Holdout Accuracy: 0.860 (86.0%) evaluated purely on predicted masks without human ground truth.

---

How to Download and Use from Hugging Face

1. Download Model Weights via Python

from huggingface_hub import snapshot_download

# Download model repository
model_path = snapshot_download(repo_id="Makifkaradag/cardiac-segmentation-diagnosis")
print("Model files downloaded to:", model_path)