TLDR: A new research paper introduces a patient-specific reinforcement learning method to create accurate 3D activity maps from just two 2D planar images in nuclear medicine. This AI-driven approach, particularly effective with diffusion models and patient-specific training, offers a faster and more cost-effective alternative to traditional 3D SPECT scans for dosimetry, leveraging pre-therapy patient data to significantly improve 3D quantification and organ delineation.
In the field of nuclear medicine, accurately mapping the 3D distribution of radiopharmaceuticals within a patient’s body is crucial for effective dosimetry, which helps determine the precise radiation dose delivered to tissues. Traditionally, this has been a challenge, often requiring expensive and time-consuming 3D SPECT (Single-Photon Emission Computed Tomography) acquisitions. While 2D planar scintigraphy offers a faster alternative, its two-dimensional nature limits its quantitative utility for 3D analysis.
A recent research paper, titled “Patient-specific AI for generation of 3D dosimetry imaging from two 2D-planar measurements,” introduces a groundbreaking methodology that could change this paradigm. Authored by Alejandro Lopez-Montes, Robert Seifert, Astrid Delker, Guido Boening, Jiahui Wang, Christoph Clement, Ali Afshar-Oromieh, Axel Rominger, and Kuangyu Shi, the study explores the use of patient-specific reinforcement learning to generate detailed 3D activity maps from just two 2D planar images (anterior and posterior views).
The core problem lies in the inherent difficulty of reconstructing a 3D image from only two 2D projections using conventional methods. The researchers tackled this by leveraging pre-therapy information, specifically PET/CT scans, to create patient-specific datasets. This involves generating various possible 3D uptake maps of radiopharmaceuticals within an individual’s anatomy, with variations in organ uptake, rotations, and translations.
Two main artificial intelligence approaches were explored: a supervised learning method using a 3DUnet architecture and an unsupervised generative model based on diffusion models. Both were trained on these patient-specific datasets. The diffusion model, in particular, was designed to generate axial slices representing possible uptakes, which were then stacked to form 3D volumes. A key aspect of the diffusion model was its ability to enforce consistency with the actual measured 2D planar images during the generation process.
The findings demonstrated a significant improvement when using patient-specific reinforcement learning compared to training networks on generic population datasets. In simulations, the mean absolute error (MAE) was reduced, and the structural similarity index measure (SSIM) increased, indicating a closer match to the ground truth 3D activity. Specifically, the diffusion model combined with patient-specific training achieved the best performance, yielding an SSIM of 0.89 in simulations and 0.73 when compared to a SPECT acquisition performed shortly after planar imaging in real cases.
Beyond quantitative metrics, the patient-specific diffusion MBIR approach also showed better delineation of organs and lesions, which is critical for accurate dosimetry. This suggests that the AI models, when trained with an understanding of an individual patient’s anatomy and expected uptake patterns, can produce more realistic and accurate 3D reconstructions.
This innovative methodology holds immense promise for nuclear medicine dosimetry. By enabling 3D quantification using only fast and less expensive planar scintigraphy, it could eliminate the need for costly and time-consuming SPECT scans. This not only reduces the burden on healthcare resources but also facilitates multi-time point dosimetry studies, which are often impractical with current 3D imaging techniques. The full research paper can be accessed here.
Also Read:
- Vision Language Models Advance Human Activity Recognition in Healthcare
- From Sketch to Structure: Automating 3D BIM Creation with Human-AI Collaboration
Ongoing work aims to further validate the method with more real-world cases, focus on lesion identification, and refine the models by explicitly incorporating pre-treatment PET/CT information during sampling. The researchers believe this approach represents a significant step towards a new era of nuclear medicine dosimetry, making 3D quantification more accessible and efficient.
