Department of Electronic and Computer Engineering, Universidad de Córdoba, Spain.
Department of Computer Science, ETH Zurich, Switzerland.
Comput Biol Med. 2020 Sep;124:103930. doi: 10.1016/j.compbiomed.2020.103930. Epub 2020 Jul 29.
Accurate and fast liver segmentation remains a challenging and important task for clinicians. Segmentation algorithms are slow and inaccurate due to noise and low quality images in computed tomography (CT) abdominal scans. Chan-Vese is an active contour based powerful and flexible method for image segmentation due to superior noise robustness. However, it is quite slow due to time-consuming partial differential equations, especially for large medical datasets. This can pose a problem for a real-time implementation of liver segmentation and hence, an efficient parallel implementation is highly desirable. Another important aspect is the contrast of CT liver images. Liver slices are sometimes very low in contrast which reduces the overall quality of liver segmentation. Hence, we implement cross-modality guided liver contrast enhancement as a pre-processing step to liver segmentation. GPU implementation of Chan-Vese improves average speedup by 99.811 (± 7.65) times and 14.647 (± 1.155) times with and without enhancement respectively in comparison with the CPU. Average dice, sensitivity and accuracy of liver segmentation are 0.656, 0.816 and 0.822 respectively on the original liver images and 0.877, 0.964 and 0.956 respectively on the enhanced liver images improving the overall quality of liver segmentation.
准确快速的肝脏分割仍然是临床医生面临的一项具有挑战性和重要的任务。由于 CT 腹部扫描中的噪声和低质量图像，分割算法既缓慢又不准确。Chan-Vese 是一种基于活动轮廓的强大而灵活的图像分割方法，因为它具有出色的抗噪能力。然而，由于耗时的偏微分方程，它的速度相当慢，特别是对于大型医疗数据集。这可能会对肝脏分割的实时实现造成问题，因此，高效的并行实现是非常需要的。另一个重要方面是 CT 肝脏图像的对比度。肝脏切片的对比度有时非常低，这降低了肝脏分割的整体质量。因此，我们实施跨模态引导的肝脏对比度增强作为肝脏分割的预处理步骤。与 CPU 相比，GPU 实现的 Chan-Vese 在增强和不增强的情况下分别将平均加速提高了 99.811（±7.65）倍和 14.647（±1.155）倍。在原始肝脏图像上，肝脏分割的平均 Dice、灵敏度和准确性分别为 0.656、0.816 和 0.822，在增强后的肝脏图像上分别为 0.877、0.964 和 0.956，提高了肝脏分割的整体质量。
