Knowledge Distillation Based Multi‐Task Learning Framework for <scp>COVID</scp> ‐19 Severity Detection

ABSTRACT Early detection of COVID‐19 is crucial for effective clinical treatment, monitoring patients, and assessing the effectiveness of treatment. Furthermore, existing approaches struggle to analyze the severity of COVID‐19 in a short period of time. Thus, the proposed study intends to use knowledge distillation‐based multi‐task learning to assess the severity of COVID‐19. Pre‐processing is first performed to remove noises and improve the quality of input images using Extended Median Filtering (Ex‐MF) from the input COVID‐CT dataset. These pre‐processed images are used as input for the proposed knowledge distillation‐based Multi‐task Learning Framework. The proposed framework includes both teacher and student networks. In the teacher network, the Residual U‐Net‐based Position Attentional Convolutional Cascaded Capsule Network (RU_PAC 3 N) is used. To reduce feature dimensionality issues, a position attention mechanism assigns weight values to each feature to achieve the essential feature subset. Finally, a cascaded capsule network is used to classify input samples. The Residual U‐Net is responsible for segmenting the affected regions from the input CT images. The teacher network is referred to as the pre‐training network. In this study, the student network uses a unique Optimized Recurrent Twin Long Short Term Memory (OR‐TLSTM) model to classify the severity of COVID‐19 disease. Furthermore, the proposed network's efficiency is improved by lowering the loss function by optimizing the weight parameters using the Binary Kookaburra Optimization (BKO) approach. The proposed detection model is optimized using an adaptive chameleon optimization (ACO) technique. The proposed method is executed using the Python tool. The proposed method achieved a better performance rate and lower error rate with 98.6% accuracy in COVID‐19 detection.