ABSTRACT
The abrupt rise in Coronavirus cases has led to shortage of rapid and highly sensitive reverse transcriptase polymerase chain reaction (RT-PCR) testing kits for the diagnosis of coronavirus disease 2019 (COVID-19). Radiologists have found X-ray images could be useful for diagnosis of COVID. In this work, Diagnostic Decision Support for Medical Imaging (DDSM)++ is introduced to detect the different abnormal conditions in lung including COVID. The scarcity of COVID dataset is handled by using various spatial transform augmentation techniques, such as power law transformation, Gaussian blur, and sharpening. Also, to get the benefit of inference accelerators, an android mobile application is developed which is quantized and optimized for ARM Mali GPU. The DDSM++ model is an extended version of DDSM model (inspired from Densenet-121), and the X-ray images are preprocessed with Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve the quality of X-ray images. The COVID X-ray images are obtained from the open source and the proposed method has obtained almost 98.47% accuracy for COVID detection. Further, the model is quantized to FP-16 using TFLITE and is utilized to benchmark the inference acceleration on Edge devices with ARM Mali GPU. About 30% and 80% reduction in inference time was observed for FP-32 and FP-16 models when run on ARM Mali GPU. Post quantization, about 5% drop in accuracy is observed for COVID detection. © 2021 IEEE.