Efficient Multi-Modal Whole Heart Segmentation via Cascaded U-Net: A Practical Solution for Clinical Settings

Cardiovascular analysis based on cardiac imaging highly benefits from automated segmentation methods. The use of deep learning is currently the most effective approach. However, its utilization in medical settings is frequently constrained by the unavailability of high-capacity hardware resources, and the high variability of medical images, which challenges the generalization ability of deep learning techniques. We propose a pipeline of two sequential U-Net for CT and MRI segmentation, configured with low complexity, allowing for usability in clinical practice. In the first stage, single-label segmentation is used to crop the image volume to a bounding box surrounding the heart. The second stage focuses on the detected region of interest. Multi-label segmentation is performed on the trimmed volume to extract 7 different substructures of the heart. Results from the WHS++ challenge validation phase show that our method achieves an average Dice Similarity Coefficient of 0.9311 on CT, and 0.8652 on MRI data. Importantly, the inference times are kept to a minimum, even when using CPU computing (∼7 s).

keywords: Convolutional Neural Network, cardiac segmentation, whole heart segmentation, CT, MRI