Convolution-Based Architecture Applied to Cardiac MRI Segmentation

Introduction

Improve efficiency and reliability in cardiac MRI segmentation by conducting a thorough comparison of the performance of various models.

Methods

• Dataset: Automated Cardiac Diagnosis Challenge (ACDC) in MICCAI challenge.
• Setup: ConvDeconv, U-Net, and GridNet.
• Evaluation: Dice Score and Hausdorff Distance.

Results

Model	LV (Dice)	LV (HD)	RV (Dice)	RV (HD)	MYO (Dice)	MYO (HD)
ConvDeconv	0.90 ± 0.020	7.15 ± 4.47	0.74 ± 0.0039	30.83 ± 10.84	30.83 ± 10.84	7.31 ± 4.46
U-Net	0.92 ± 0.019	9.61 ± 5.83	0.78 ± 0.0038	25.50 ± 9.72	0.87 ± 0.012	6.07 ± 3.66
GridNet	0.92 ± 0.018	6.86 ± 4.56	0.80 ± 0.0037	16.02 ± 7.24	0.87 ± 0.013	4.90 ± 2.63