FPGA Oriented Lightweight Deep Learning Inference for Liver Cancer Segmentation

Field-programmable gate arrays (FPGAs) have succeeded in deep learning (DL) inference due to their high parallelism and low power consumption, potentially enabling the miniaturization of DL-based computer aided diagnosis (CAD) system using liver cancer segmentation. However, state-of-the-art deep learning backbone models tend to enlarge their model capacity or design complex unfolding architectures, which may hinder the implementation on FPGA with relatively slow updates of specs and computational resources. This paper proposes a new method for lightweight inference on FPGAs that combines lightweight modeling and FPGA-based model compression. Firstly, our proposed split-attention-shuffle bottleneck (SASB) is the module that could provide fast adaptation of light weighting to a standard network simply through layer substitution. Secondly, we employ the model compression techniques by means of standard quantization and compilation methods on an in-house prototyping edge device named HIDE, significantly reducing precision of inference model weights but preserving the model accuracy. Experiments show that our method saves budgets of computational resources while preserving comparable performance compared to other heavy-weight models. Our approach is not limited to specialized edge computing devices and can be extended to other deep learning methods and medical imaging modalities, enhancing the potential and impact of portable CAD systems.