Abstract
Lung ultrasound images have shown great promise to be an operative point-of-care test for the diagnosis of COVID-19 because of the ease of procedure with negligible individual protection equipment, together with relaxed disinfection. Deep learning (DL) is a robust tool for modeling infection patterns from medical images; however, the existing COVID-19 detection models are complex and thereby are hard to deploy in frequently used mobile platforms in point-of-care testing. Moreover, most of the COVID-19 detection models in the existing literature on DL are implemented as a black box, hence, they are hard to be interpreted or trusted by the healthcare community. This paper presents a novel interpretable DL framework discriminating COVID-19 infection from other cases of pneumonia and normal cases using ultrasound data of patients. In the proposed framework, novel transformer modules are introduced to model the pathological information from ultrasound frames using an improved window-based multi-head self-attention layer. A convolutional patching module is introduced to transform input frames into latent space rather than partitioning input into patches. A weighted pooling module is presented to score the embeddings of the disease representations obtained from the transformer modules to attend to information that is most valuable for the screening decision. Experimental analysis of the public three-class lung ultrasound dataset (PCUS dataset) demonstrates the discriminative power (Accuracy: 93.4%, F1-score: 93.1%, AUC: 97.5%) of the proposed solution overcoming the competing approaches while maintaining low complexity. The proposed model obtained very promising results in comparison with the rival models. More importantly, it gives explainable outputs therefore, it can serve as a candidate tool for empowering the sustainable diagnosis of COVID-19-like diseases in smart healthcare.