High Resolution Medical Image Segmentation Using Data-Swapping Method

Abstract

Deep neural network models used for medical image segmentation are large because they are trained with high-resolution three-dimensional (3D) images. Graphics processing units (GPUs) are widely used to accelerate the trainings. However, the memory on a GPU is not large enough to train the models. A popular approach to tackling this problem is patch-based method, which divides a large image into small patches and trains the models with these small patches. However, this method would degrade the segmentation quality if a target object spans multiple patches. In this paper, we propose a novel approach for 3D medical image segmentation that utilizes the data-swapping, which swaps out intermediate data from GPU memory to CPU memory to enlarge the effective GPU memory size, for training high-resolution 3D medical images without patching. We carefully tuned parameters in the data-swapping method to obtain the best training performance for 3D U-Net, a widely used deep neural network model for medical image segmentation. We applied our tuning to train 3D U-Net with fullsize images of 192 × 192 × 192 voxels in brain tumor dataset. As a result, communication overhead, which is the most important issue, was reduced by 17.1%. Compared with the patch-based method for patches of 128 × 128 × 128 voxels, our training for full-size images achieved improvement on the mean Dice score by 4.48% and 5.32 % for detecting whole tumor sub-region and tumor core subregion, respectively. The total training time was reduced from 164 hours to 47 hours, resulting in 3.53 times of acceleration.

Generation of 3D Tumor Models from DICOM Images for Virtual Planning of its Recession

“…[37], Maitra et al [38] and Zhang et al [39], and the one in this paper lays mainly in the methodology for presenting results, where, in addition to segmentation, a threedimensional reconstruction algorithm is incorporated. On the other hand, solutions presented by Zhang et al [40], Zhao et al [41], Shen et al [42], Holger et al [43] show the data using three-dimensional reconstruction.…”

Multi-stage learning for segmentation of aortic dissections using a prior aortic anatomy simplification

“…The inherent shortcoming of using DL-based methods in medical 3-D scenes (Livne et al, 2019;López-Linares et al, 2018) were the large memory requirements and the resultant hard-to-deploy challenges of the models in the practice (Imai et al, 2018). In CT images, AO and BR were distributed in all axial slices, so the entire CT image should be used for the network training, making the memory requirements even more severe.…”

Accelerating 3D Medical Image Segmentation by Adaptive Small-Scale Target Localization

“…Several researchers have recently studied ways to accelerate CNNs for 3D medical imaging segmentation. The authors of [ 20 ] proposed a CPU-GPU data swapping approach that allows for training a neural network on the full-size images instead of the patches. Consequently, their approach reduces the number of iterations on training, hence it reduces the total training time.…”