LAUSR

An automatic tool, targeting low-cost, low-power, point-of-care embedded system, is proposed for fluorescence diagnostic imaging. This allows for a quick and accurate diagnosis even when used by nonexpert operators. To achieve this goal, an embedded system has been equipped with an end-to-end deep-learning algorithm that does not require manual parameter tuning to perform a diagnosis. The proposed deep convolutional model, named BrightNet, is based on a single-shot detector neural network, modified to estimate the brightness of the detected fluorescent spots in a low-density protein or DNA microarray and finalize the diagnosis. Several optimization steps are presented to compress the inference model size, which is required for the deployment into a portable resource-constrained device. The resulting inference time is about 66 [ms] on an i7 3770K desktop CPU and is estimated to be lower than 5 [s] on an ARM-Cortex M7 considering $1.1 \times 10^{9}$ multiply-accumulate operations. BrightNet has been successfully validated for the detection and discrimination of four different serotypes of the dengue virus in a set of human samples as well as for the diagnosis of West Nile virus in horse sera. When evaluated on the considered diagnostic tasks, BrightNet provides better average accuracy than a state-of-the-art variational approach that requires operator intervention, with significant additional advantages of complete automation and quicker diagnosis.