DPHT-Net: A Novel Self-Supervised Hybrid CNN-Transformer Approach for Automated Pulmonary Embolism Classification in CT Pulmonary Angiogram Scans

Abstract

Pulmonary embolism (PE) is a life-threatening condition that requires timely and accurate diagnosis to enable appropriate treatment and reduce mortality rates. Despite significant advances in medical imaging, reliable detection of PE in computed tomography pulmonary angiography (CTPA) remains challenging due to the subtle appearance of emboli, variations in contrast, and the high-dimensional nature of volumetric data. These challenges necessitate robust and efficient automated methods to support clinical decision-making. In this study, we propose a diffusion-pretrained hybrid Convolutional Neural Network (CNN)-Transformer network (DPHT-Net), a lightweight architecture designed to effectively capture both local and global embolic patterns in CTPA scans. The proposed framework integrates self-supervised diffusion-based pretraining, a CNN-based module for local feature extraction and refinement, and a Transformer-based component for modeling long-range inter-slice dependencies. In addition, a sinh–cosh-based preprocessing step is introduced to enhance image contrast and highlight subtle embolic regions. The proposed model was evaluated on the RSNA-STR PE dataset. Experimental results demonstrate that DPHT-Net achieves an accuracy of 96.4% and an F1-score of 93.8%, outperforming conventional CNN-based methods by 11.9% in accuracy and 12.7% in F1-score, and surpassing Transformer-based approaches by 5.4% and 4.2%, respectively. These results indicate that DPHT-Net provides a robust, computationally efficient, and clinically applicable solution for automated PE detection, offering a promising direction for volumetric medical image analysis and computer-aided diagnosis systems.