Padm: Physics-aware Diffusion Model Corrects SPECT Attenuation Without CT, Enabling Wider Cardiac Imaging Access

Attenuation errors continue to plague cardiac Myocardial Perfusion Imaging using Single-Emission Tomography, frequently reducing diagnostic confidence and hindering clear clinical interpretation. Trung Kien Pham, Hoang Minh Vu, and Anh Duc Chu, working with colleagues at their institutions, address this challenge with a novel, cost-effective solution that eliminates the need for expensive and potentially harmful Computed Tomography scans. The team introduces the Physics-aware Attenuation Correction Diffusion Model, or PADM, a generative method that leverages fundamental physics principles to correct these errors using only the raw SPECT data. Supporting this breakthrough, they also present CardiAC, a comprehensive dataset of over 400 patient studies, which facilitates robust training and validation of the new technique, demonstrably outperforming existing methods in both quantitative accuracy and visual quality.

While hybrid SPECT/CT systems can correct for this, their cost and the associated radiation exposure limit their widespread use. Researchers are now developing innovative, CT-free solutions to improve diagnostic accuracy in cardiac SPECT imaging.

Deep Learning Enables CT-Free SPECT Correction

This research details advancements in medical image processing, specifically focusing on attenuation correction for SPECT imaging in cardiac applications. The core challenge lies in accurately correcting for photon attenuation without relying on costly or radiation-emitting CT scans. Current approaches explore the use of deep learning techniques to achieve this goal, including Generative Adversarial Networks (GANs) and, increasingly, diffusion models. Diffusion models are emerging as a powerful alternative to GANs, often achieving higher image quality and more stable training. Some methods aim to directly correct the SPECT image itself, bypassing the need to estimate an attenuation map. The team trained PADM using a teacher-student distillation mechanism, enabling accurate attenuation correction from standard, non-corrected SPECT data. To support this research and provide a valuable resource for the scientific community, the team created CardiAC, a comprehensive dataset comprising 424 patient studies.

This unique dataset provides paired non-corrected and corrected SPECT reconstructions alongside high-resolution CT-based attenuation maps, establishing a strong benchmark for future investigations. The high-resolution volumes within CardiAC offer detailed anatomical information under both stress and rest conditions. Experiments demonstrate that PADM consistently outperforms existing generative models, validated through quantitative metrics and visual assessment. This confirms the effectiveness of integrating physical principles with advanced diffusion modeling, delivering a substantial advancement in cardiac SPECT imaging and potentially broadening access to this important diagnostic tool while reducing patient exposure to radiation. PADM utilizes a diffusion-based generative approach, incorporating explicit physics-based knowledge through a teacher-student learning framework, to reconstruct accurate images from standard, non-attenuation-corrected SPECT data. To facilitate this work and provide a benchmark for future research, the team introduced CardiAC, a comprehensive dataset containing paired non-attenuation-corrected and attenuation-corrected SPECT reconstructions, alongside high-resolution CT-derived attenuation maps.

Extensive evaluations demonstrate that PADM consistently outperforms existing generative models, delivering superior reconstruction fidelity in both quantitative metrics and visual assessments. This improvement is particularly noticeable in clinically important regions of the myocardium, suggesting potential for enhanced diagnostic accuracy. Future research will focus on collaboration with clinicians to evaluate reconstructed images in real-world clinical scenarios, including lesion classification and automated report generation, building on recent advances in medical image analysis and multimodal vision-language modeling.