Impact of Attention Modules in Deep Learning-based Semantic Segmentation: Evaluation for Liver Lesion Segmentation from CT Images

Abstract

Given the potential performance of deep learning-based systems, there are various approaches for medical image analysis, particularly attention-based semantic segmentation models. Attention-based techniques will facilitate more accurate predictions in medical image segmentation by concentrating on the correct region. The purpose of this study is to analyze multiple attention-based deep learning methods for semantic segmentation tasks in the medical imaging field. Since the attention mechanism was introduced, independent of its structure, it has been progressively applied to enhance the performance of deep learning systems in a variety of medical image processing applications. In addition, different deep learning networks for liver and tumor segmentation have been proposed utilizing an attention mechanism. Three deep learning models for liver and tumor segmentation were evaluated in this study to provide a baseline for comparing the performance of the various models. A total of 131 computed tomography (CT) volumes were included in this investigation (104 subjects for training and 27 subjects for validation). The segmented images with varying degrees of tumor contrast were evaluated using a variety of different evaluation metrics. The performance of implemented models was evaluated on LiTS 2017 Challenge dataset. Compared to two other implemented models, the attention-based U-Net with 2.32 million parameters was more accurate. The liver and tumor segmentation dice coefficients in the best model were 0.934 and 0.778, respectively. Experimental results on the LiTS datasets showed that the attention mechanism is potentially capable of producing improved performance. © 2022 IEEE.