Training Models 20X Faster in Medical Image Analysis

Dong Yang, Holger Roth, Xiaosong Wang, Ziyue Xu, Yan Cheng, Daguang Xu

Keywords: Medical image analysis, deep learning, segmentation.

Track: full conference paper

Paper Type: both

Abstract: Analyzing high-dimensional medical images (2D/3D/4D CT, MRI, histopathological images, etc.) plays an important role in many biomedical applications, such as anatomical pattern understanding, disease diagnosis, and treatment planning. The AI assisted models have been widely adopted in the domain of medical image analysis with great successes. However, training such models with large-size data is expensive in terms of computation and memory consumption. In this work, we provide solutions for improving model training efficiency, which will speed up the training of AI models (20X faster on an exemplary 3D segmentation framework), and enable researchers and radiologists to improve the efficiency in their clinical studies. The overall efficiency improvement comes from both improved algorithms and engineering advance.