Two Deep Learning Approaches for Automated Segmentation of Left Ventricle in Cine Cardiac MRI

TL;DR

This work tackles automated left-ventricle segmentation in cine cardiac MRI by introducing two LN- and BI-based U-Net variants, LNU-Net and IBU-Net, and systematically comparing normalization strategies (BatchNorm, LayerNorm, Batch-InstanceNorm). Leveraging an encoder–decoder FCN with skip connections and ELU activations, along with data augmentation, the study demonstrates that Batch-Instance Normalization combined with ELU delivers the highest $Dice$ mean (~0.96) and lowest $APD$ (~1.9 mm), while also accelerating training. The best-performing model, IBU-Net with augmentation, outperforms the baseline U-Net and other variants across metrics. These results suggest substantial gains in segmentation accuracy and training efficiency, with direct implications for clinical LV function assessment in cine MRI workflows.

Abstract

Left ventricle (LV) segmentation is critical for clinical quantification and diagnosis of cardiac images. In this work, we propose two novel deep learning architectures called LNU-Net and IBU-Net for left ventricle segmentation from short-axis cine MRI images. LNU-Net is derived from layer normalization (LN) U-Net architecture, while IBU-Net is derived from the instance-batch normalized (IB) U-Net for medical image segmentation. The architectures of LNU-Net and IBU-Net have a down-sampling path for feature extraction and an up-sampling path for precise localization. We use the original U-Net as the basic segmentation approach and compared it with our proposed architectures. Both LNU-Net and IBU-Net have left ventricle segmentation methods: LNU-Net applies layer normalization in each convolutional block, while IBU-Net incorporates instance and batch normalization together in the first convolutional block and passes its result to the next layer. Our method incorporates affine transformations and elastic deformations for image data processing. Our dataset that contains 805 MRI images regarding the left ventricle from 45 patients is used for evaluation. We experimentally evaluate the results of the proposed approaches outperforming the dice coefficient and the average perpendicular distance than other state-of-the-art approaches.

Figures (5)

• Figure 1: Architecture of the BNU-Net convolutional network. (a) The contraction path is responsible for feature extraction. (b) Batch normalization is performed after each convolution in the convolutional layer.
• Figure 2: The LNU-Net architecture of the proposed fully convolutional network. Layer normalization is performed after each convolution in the convolutional layer.
• Figure 3: Architecture of the IBU-Net convolutional network. Instance normalization is applied in the first convolutional layer. Batch normalization is performed after each convolution in the convolutional layer.
• Figure 4: Examples of segmentation results on raw inputs from three conditions in Sunnybrook dataset. The first row contains heart failure with infarction, the second row represents hypertrophy, the third row shows healthy patients.
• Figure 5: Some segmentation outputs by our methods. The solid lines represent the segmentation examples of Sunnybrook dataset. We compare four different network methods, which are U-Net, BNU-Net, LNU-Net, and IBU-Net respectively.