On How to Push Efficient Medical Semantic Segmentation to the Edge: the SENECA approach

Semantic segmentation is the process of assigning each input image pixel a value representing a class, and it enables the clustering of pixels into object instances. It is a highly employed computer vision task in various fields such as autonomous driving and medical image analysis. In particular, in medical practice, semantic segmentation identifies different regions of interest within an image, like different organs or anomalies such as tumors. Fully Convolutional Networks (FCNs) have been employed to solve semantic segmentation in different fields and found their way in the medical one. In this context, the low contrast among semantically different areas, the constraint related to energy consumption, and computation resource availability increase the complexity and limit their adoption in daily practice. Based on these considerations, we propose SENECA to bring medical semantic segmentation to the edge with high energy efficiency and low segmentation time while preserving the accuracy. We reached a throughput of 335.4 +/- 0.34 frames per second on the FPGA, 4.65 x better than its GPU counterpart, with a global dice score of 93.04% +/- 0.07 and an improvement in terms of energy efficiency with respect to the GPU of 12.7x.