Enhancing Efficiency in Cardiac 3D Modelling through Application, Adaptation, and Analysis of a 3D U-Net architecture-based Computed Tomography Image Segmentation Algorithm

Aim and Research Question(s)

This thesis aims to improve the efficiency and accuracy of cardiac 3D modelling by implementing an AI-based segmentation algorithm, using 3D U-Net architecture, for cardiac CT images. To achieve this goal my research question was the following: "How does the integration of an U-Net algorithm into cardiac CT image segmentation affect the efficiency and accuracy of results compared with manual segmentation?"

Background

Accurately segmenting cardiac structures in CT images is crucial for creating detailed 3D visualizations used in clinical decisions [1]. Recent advancements in machine learning, especially convolutional neural networks (CNNs), have transformed medical image analysis. The U-Net architecture, a type of CNN, excels in image segmentation by capturing both local and global features [2]. However, as Alnasser et al. [3] suggests, U-Net performance can vary significantly based on the dataset and implementation, making systematic evaluations essential for understanding the strengths and limitations of deep learning models in cardiac image segmentation.

Methods

In this thesis, three segmentation methods for the left ventricle (LV) in cardiac contrast-enhanced CT (CCTA) images were compared: manual segmentation (gold standard), a MONAI-based 3D U-Net model, and the commercial AI software Mimics V26. A total of 50 anonymized CCTA datasets were retrospectively obtained from "Mein Hanusch Hospital" and divided into a training set (N = 47) and a test set (N = 3). Each training dataset was manually labeled using 3D Slicer, while the test sets were segmented manually by an experienced radiographer for later comparison with the AI-based segmentations. The MONAI-based model was trained using the 47 training datasets, and its performance was evaluated against the manual and Mimics segmentations. Metrics such as Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and LV volumetric measurements were used to assess segmentation accuracy. Efficiency was evaluated based on segmentation time, and two experts (a radiologist and a cardiologist) conducted a blinded assessment of the quality of the segmentations.

Results and Discussion

The MONAI-based model achieved DSC values between 0.88 and 0.91. Compared to manual segmentation, LV volumes were consistently under-segmented, while Mimics showed over-segmentation. The average Hausdorff distance for Manual vs. MONAI and Manual vs. Mimics was 1.45 mm ± 0.23 mm and 0.52 mm ± 0.25 mm, respectively. Experts noted artefacts and incomplete segmentations with MONAI-based AI, resulting in lower ratings compared to Mimics, which demonstrated higher accuracy. However, AI methods significantly reduced segmentation time. 

Conclusion

Self-trained AI models using MONAI offer an alternative to commercial tools like Mimics. Despite promising results, qualitative evaluations highlight the need for rigorous validation and standardization to ensure clinical reliability. Close collaboration with medical professionals is essential for integrating AI into clinical practice.

References

_{[1] Y. Song, S. Ren, Y. Lu, X. Fu, and K. K. L. Wong, “Deep learning-based automatic segmentation of images in cardiac radiography: A promising challenge,” Computer Methods and Programs in Biomedicine, vol. 220, p. 106821, Jun. 2022, doi: 10.1016/j.cmpb.2022.106821. [2] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation.” arXiv, May 18, 2015. doi: 10.48550/arXiv.1505.04597. [3] T. N. Alnasser et al., “Advancements in cardiac structures segmentation: a comprehensive systematic review of deep learning in CT imaging,” Front. Cardiovasc. Med., vol. 11, Jan. 2024, doi: 10.3389/fcvm.2024.1323461.}