Data-Type Assessment for Real-Time Hyperspectral Classification in Medical Imaging

Real-time constraints in image processing applications often force their optimization using hardware accelerators. This is the case for intraoperative medical images used during surgical procedures. In this context, the challenge consists in processing large volumes of data while employing high complexity algorithms in a limited period of time. Newly developed algorithms must meet both quality-accurate and hardware-efficient characteristics. In this work, we have evaluated the impact of using different data types in a processing chain to classify tissues from hyperspectral video in surgical environments. The software was run on two different embedded CPU+GPU platforms. The results show an improvement in performance by up to 9 times without increasing power consumption by reducing the bit depth from 64 to 16. The impact these reduction have on quality has been measured analytically, by calculating the RMSE, and subjectively, by surveying neurosurgeons. In both cases the results show a minimal impact on the overall quality.