Simultaneous segmentation and classification of lung CT scans for COVID-19 diagnosis: a deep multi-task learning perspective

• Published in: Neural computing & applications Vol. 37; no. 6; pp. 4185 - 4205
• Main Authors: Kordnoori, Shirin, Sabeti, Maliheh, Mostafaei, Hamidreza, Banihashemi, Saeed Seyed Agha
• Format: Journal Article
• Language: English
• Published: London Springer London 01.02.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Classification; Coders; Computational Biology/Bioinformatics; Computational Science and Engineering; Computed tomography; Computer Science; Conditional random fields; COVID-19; Data Mining and Knowledge Discovery; Image Processing and Computer Vision; Image segmentation; Lungs; Medical imaging; Original Article; Preprocessing; Probability and Statistics in Computer Science; COVID-19; Segmentation; Multi-task learning; Classification
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-024-10809-8

In response to the pressing need for intelligent systems to identify COVID-19 cases and detect infected areas in lung CT scan images during the ongoing pandemic, there is a growing demand for integrated solutions that can effectively handle both segmentation and classification tasks simultaneously. Such systems can provide valuable support to clinicians while enhancing diagnostic accuracy by leveraging shared features. This paper introduces a multi-task architecture designed to concurrently process lung CT scan images for both segmentation and classification tasks. The proposed approach initially employs a shared encoder–decoder architecture based on U-Net, augmented by an additional branch dedicated to classification using a perceptron. To address performance disparities between these tasks during preprocessing, a combination of binary preprocessing algorithms is introduced to establish task equilibrium. Furthermore, convolution block attention module is incorporated into the encoder levels to enhance inter-task coherence. Additionally, the application of conditional random field serves for post-processing in the segmentation task. The effectiveness of the proposed structure is evaluated across four datasets, demonstrating its superiority over previous studies and pre-trained models. The results showcase the highest classification accuracy at 98.14 ± 1.00, accompanied by a Dice index of 89.91 ± 2.00 in segmentation, alongside other relevant evaluation metrics. Furthermore, the multi-task framework is successfully applied to U-Net++ and ResUNet architectures, yielding favorable outcomes. Our proposed model provides a robust solution for accurate COVID-19 analysis, offering valuable support to medical professionals in making diagnostic and treatment decisions.