| Summary: | Medical image segmentation plays a vital role in automated disease screening and diagnosing systems. One of the diseases that depend on image segmentation output is COVID-19, in which the diseased lesion is informative in order to determine the disease severity. The task of segmenting COVID-19 lesions is a challenging process, not only due to computational costs and hardware limitations but also because of the varying sizes and shapes of the lesions. In this study, we proposed to integrate a multiscale unit, specifically the ASPP, into a base DABNet, resulting in a lightweight module called DAASPP. The addition of a multiscale unit is meant to further improve the network capability in extracting lesion features of various scales so that small-and large-scale lesions can be segmented accurately. There are two variations of DAASPP have been proposed: one with a single-unit ASPP and another network with a dual-unit ASPP. Both variations are placed at the last and second last DABNet blocks due to the availability of complex features at these two layers. Four different scales are explored for each network variation, specifically the 2, 3, 4, and 5 dilated convolution layers. The results showed that DAASPP with a single-unit ASPP and four parallel scale branches demonstrated the best segmentation performance, achieving mean IoU, mean Dice coefficient, and mean accuracy scores of 0.6629, 0.7461, and 0.9991, respectively. This improvement in extracting the lesion regions will help in producing a better diagnosis system of the COVID-19 severity level. © 2024 IEEE.
|
|---|