Multiplanar Data Augmentation and Lightweight Skip Connection Design for Deep-Learning-Based Abdominal CT Image Segmentation

In recent years, deep-learning-based computed tomography (CT) image segmentation methods show state-of-the-art performance. However, applying deep learning in CT image analysis is challenging because well-annotated medical data are expensive and time-consuming. This article proposes two steps to improve the abdominal CT image segmentation performance on top of the same labeled dataset. The innovations are twofold. First, there are a number of medical image data augmentation methods in the literature, but the volumetric measurement of CT images is not well-emphasized. In this article, a new 2.5-D CT image augmentation analysis is presented to alleviate the aforementioned issue from both the theoretical and experimental perspectives. The theory derived in this article provides a rationale first time in the literature for training a single model using very different CT images. The theory is verified by three 2.5-D CT image augmentation methods and their seven combinations to make full use of the existing dataset. Second, skip connections can greatly affect the performance of encoder-decoder networks, which have shown effectiveness in extracting fine-grained features of target objects. This article redesigns three skip connections, from complex to simple, based on the original TransUNet to extract features of different semantic scales and aggregate them at the decoder for more flexible feature fusion. The results show that a simple skip connection design may achieve better performance. Extensive experiments are conducted to verify the proposed data augmentation and skip connection designs and compare with some selected state-of-the-art methods. Experimental results show that the proposed new lightweight skip connection (LSC) together with the proposed data augmentation methods can greatly improve the performance without increasing the new labeled data.