Phonocardiogram Segmentation with Tiny Computing

Abstract

The stethoscope is a daily used tool that allows medical doctors to diagnose common cardiovascular diseases by listening to heart sounds. However, dedicated medical training is required to operate it. Numerous machine learning techniques have been used in attempts to automate this process and have yielded highly accurate results. However, creating a low power, portable, economical, and accurate machine learning stethoscope calls for tiny processing of phonocardiograms i.e., heart sound digital processing to run within an embedded device. To address the need to deploy the solution within a constrained tiny device, we propose an 8-bit deep learning model with low embedded FLASH and RAM utilization of 126 KiB and 45 KiB respectively, which is optimized for inference on an off-the-shelf STM32H7 microcontroller with an inference time of 12 ms, in 126KiB FLASH and 45 KiB RAM being 91.65% accurate.

Low-Cost FPGA Implementation of Deep Learning-Based Heart Sound Segmentation for Real-Time CVDs Screening

“…Note that the three models have been trained with different datasets, so performance may depend on this factor. Finally, [20] reported a lower model inference time than this work. However, this could be related to the fact that they used a significantly smaller model with only three convolutional layers, halving the number of bits used in their implementation (8-bit representation against 16bit), and they used 4.8 times the frequency employed in this work.…”

“…Nonetheless, it shows the state-of-the-art in this field, setting the basis for further research to improve accuracy, reduce inference times or lower the power consumption. Firstly, the clock frequency of this implementation is 4.8 times lower than the one employed in [20] and 15 times lower than in [21]. This, together with the fact that FPGAs are less power demanding than CPUs would expectedly imply a significant decrease in the power consumption of this implementation compared to the other two.…”

“…Given this, it is clear that our model has significantly more complex than the other two. For this reason, the classification results reported in this work achieved a more accurate segmentation considering the four heart sound components of a PCG, whereas [20] distinguished between S1, S2 and the rest of the signal, and [21] limited the model to only systole and diastole detection. Note that the three models have been trained with different datasets, so performance may depend on this factor.…”

“…Only a few studies have presented hardware implementations of heart-sound segmentation algorithms based on deep learning. Kwiatkowski et al [20] implemented a small CNN in an ARM Cortex M7 processor with an inference time of 11 ms, using an 8-bit representation. Vakamullu et al [21] used a Raspberry Pi 3B (quad-core ARM Cortex A53) to implement a 1-dimensional CNN.…”

“…All works have used a CNN to perform the segmentation, differing in their architecture. Kwiatkowski et al [20] used three convolutional layers with intercalated maxpooling and batch-normalization layers, and Vakamullu et al [21] implemented just two convolutional layers followed by max-pooling layers. Our model has 23 convolutional layers, and uses the U-Net architecture, with encoding and decoding stages.…”