Explainable Deep Learning Framework for ground glass opacity (GGO) Segmentation from Chest CT scans

Abstract

Segmenting ground glass opacities (GGO) from chest computed tomography (CT) scans is crucial for early detection and monitoring of lung diseases. This includes lung infections and acute alveolar malignancies. However, GGO segmentation is a challenging task in chest radiology as GGOs often exhibit a range of characteristics and displays low-intensity contrast with adjacent structures in CT images. This study introduces a novel deep learning framework for segmenting GGOs in CT scans using ResNet-50U-Net, which is an improved U-Net model with a pretrained ResNet-50 to enhance feature extraction. A total 62 CT pseudoanonymised images were collected from patients with Covid-19, annotated by experienced radiologist, and further processed for analysis. Our experimental results demonstrate that the proposed ResNet-50U-Net outperforms the standard U-Net as well as DenseNet-121U-Net architectures in detecting the GGO locations with Dice similarity score, Precision, and Recall of 0.71, 0.63, and 0.83, respectively. Unlike current deep learning-enabled methods for GGO segmentation, which face trust challenges due to their "black-box" nature, our approach integrates a post-hoc visual explainability feature through the GradCAM++ (Gradient-weighted Class Activation Mapping) algorithm. This tool highlights significant regions within the Chest CT scans that impacts the model's decision, providing beneficial insights into the segmentation process.