A Lightweight Deep Learning Approach for Liver Segmentation

Liver segmentation is a prerequisite for various hepatic interventions and time-consuming manual task performed by radiology experts. Recently, computationally expensive deep learning architectures tackled this aspect without considering the resource limitations of real-life clinical setup. In paper, we investigated capabilities lightweight model, UNeXt, in comparison with U-Net model. Moreover, conduct broad analysis at micro macro levels these using two training loss functions: soft dice unified focal loss, substituting commonly used ReLU activation function, novel Funnel function. An automatic post-processing step that increases overall performance models also proposed. Model evaluation were on public database—LiTS. The results show UNeXt model (Funnel activation, step) achieved 0.9902 similarity coefficient whole CT volumes test set, 15× fewer parameters nearly 4× less inference time, compared to its counterpart, U-Net. Thus, can become new standard medical segmentation, when implemented thoroughly alleviate computational burden while preserving parameter-heavy architecture.