Integrating Dual-Supervision and Region Fusion Attention for COVID-19 Image Segmentation

Purpose: Effectively and precisely segmenting infections in COVID-19 images is crucial for automated clinical diagnostics. This study introduces a deep-learning framework for precise segmentation of COVID-19 infections in chest radiology images, leveraging lung region and the region attention to overcome the challenges in current segmentation practices.

Methods/Materials: COVID-19 datasets were curated and utilized for training and validation, comprising CT dataset (Train=1844, Test=373 slices) and X-ray dataset (Train=1864, Validation=466, Test=583 slices). The proposed deep learning network incorporates Dual-Supervision at multi-scale levels and integrates a region fusion attention module to enhance the precision of infection segmentation. Hyperparameters, including a hybrid loss function, implementation of AdamW optimizer with a learning rate of 0.0002, and a batch size of 4, were fine-tuned for optimal performance. Evaluation metrics such as Intersection over Union, Dice Coefficient, Mean Absolute Error, and F-Score were employed to compare the performance of our model against other state-of-the-art models.

Results: Our proposed network consistently outperforms other state-of-the-art models, encompassing diverse UNets, general segmentation networks, and other state-of-the-art COVID-19 segmentation models. In an ablation study designed to validate the effectiveness of Dual-Supervision and the region fusion attention module, our method surpasses the baseline, notably improving the Dice coefficient by up to 14.24% and IoU by 28.64%. The introduction of the novel region fusion attention module further refines the precision of segmentation, underscoring the efficacy of our method in medical image analysis.

Conclusions: In summary, our proposed network significantly enhances COVID-19 segmentation accuracy, advancing computer-aided diagnosis. Its targeted analysis prioritizes critical cases, potentially improving patient outcomes. Future work includes extending applicability to severity assessments and refining X-ray imaging utility. Our proposed network represents a noteworthy stride in leveraging computational methods for effective COVID-19 diagnostics.