Explore the scholarly record.

Abstract

Abstract Chest X-ray (CXR) images have been used for screening purposes in thoracic diseases owing to their low cost and wide availability. However, accurate classification of thoracic diseases using CXRs with limited data remains challenging due to poor lesion visualization and overlapping disease-sensitive structures. On the contrary, computed tomography (CT) scans provide better three-dimensional, disease-sensitive detail and more precise visualization of pulmonary lesions. This paper proposes CTGuideNet, an adaptive cross-modality knowledge distillation framework that provides CT-guided feature supervision by transferring disease-sensitive feature representations learned from independently collected CT images to chest X-ray classification models. Contrary to existing approaches to knowledge distillation (KD) where uniform distillation strategies across different architectures are used, we design an approach in which distillation behavior varies with the representational capabilities of various convolutional neural network (CNN) architectures like lightweight CNNs, residual CNNs, and attention-aware CNNs. First, a CT-based teacher model was trained on 8408 CT images of COVID-19, pneumonia, lung tumors, and normal conditions; then extensive experiments were conducted in low-data and large-scale CXR settings containing COVID-19, pneumonia lung tumors, and normal classes. The proposed framework established the largest improvements under limited-data conditions, where CT-Guided supervision improved classification performance by up to 10% over baseline training in the multi-architecture experiments. The lightweight MobileNetV2 demonstrated consistent improvements under multi-seed evaluation. Additional Grad-CAM analysis and external validation on the JSRT lung nodule dataset further confirmed improved lesion-sensitive representation learning and cross-dataset generalization capability. The experimental results demonstrate that CT-guided feature supervision using disease-sensitive representations learned from CT images provides an effective strategy for improving data-efficient thoracic disease classification in CXR images.