Light-weighted ensemble network with multilevel activation visualization for robust diagnosis of COVID19 pneumonia from large-scale chest radiographic database

Currently, the coronavirus disease 2019 (COVID19) pandemic has killed more than one million people worldwide. In the present outbreak, radiological imaging modalities such as computed tomography (CT) and X-rays are being used to diagnose this disease, particularly in the early stage. However, the assessment of radiographic images includes a subjective evaluation that is time-consuming and requires substantial clinical skills. Nevertheless, the recent evolution in artificial intelligence (AI) has further strengthened the ability of computer-aided diagnosis tools and supported medical professionals in making effective diagnostic decisions. Therefore, in this study, the strength of various AI algorithms was analyzed to diagnose COVID19 infection from large-scale radiographic datasets. Based on this analysis, a light-weighted deep network is proposed, which is the first ensemble design (based on MobileNet, ShuffleNet, and FCNet) in medical domain (particularly for COVID19 diagnosis) that encompasses the reduced number of trainable parameters (a total of 3.16 million parameters) and outperforms the various existing models. Moreover, the addition of a multilevel activation visualization layer in the proposed network further visualizes the lesion patterns as multilevel class activation maps (ML-CAMs) along with the diagnostic result (either COVID19 positive or negative). Such additional output as ML-CAMs provides a visual insight of the computer decision and may assist radiologists in validating it, particularly in uncertain situations Additionally, a novel hierarchical training procedure was adopted to perform the training of the proposed network. It proceeds the network training by the adaptive number of epochs based on the validation dataset rather than using the fixed number of epochs. The quantitative results show the better performance of the proposed training method over the conventional end-to-end training procedure.

A large collection of CT-scan and X-ray datasets (based on six publicly available datasets) was used to evaluate the performance of the proposed model and other baseline methods. The experimental results of the proposed network exhibit a promising performance in terms of diagnostic decision. An average F1 score (F1) of 94.60% and 95.94% and area under the curve (AUC) of 97.50% and 97.99% are achieved for the CT-scan and X-ray datasets, respectively. Finally, the detailed comparative analysis reveals that the proposed model outperforms the various state-of-the-art methods in terms of both quantitative and computational performance.