Task-based evaluation of a scatter projection and deep learning-based transmission-less attenuation-compensation method with clinical data

Introduction: Myocardial perfusion single-photon emission computed tomography (SPECT) (MPS) is a widely used imaging modality in the diagnosis of coronary artery disease. Accurate attenuation compensation (AC) is beneficial for clinical interpretation of MPS images (1). Conventional AC methods typically require an additional CT transmission scan, which has multiple disadvantages, including increased radiation dose, higher costs, and more importantly, potential misalignment between SPECT and CT images (2). Moreover, many SPECT systems do not contain a CT component, including those with solid-state detectors and those used in remote settings. Thus, there is a crucial need for transmission-less (tx-less) AC methods for MPS. To address this need, multiple tx-less AC methods have been proposed, including deep learning (DL)-based methods (3-7). Promising results from these methods using fidelity-based metrics motivate evaluation of performance on clinical tasks. In this context, we recently proposed a scatter projection and DL-based tx-less AC method (SLAC) for SPECT (7). Our goal in this study is to objectively evaluate the performance of this method on the task of detecting cardiac defects from MPS images using an observer study with clinical data.

Methods: The SLAC method integrates physics and DL to obtain an estimate of the attenuation map, which is used for AC (7,8). More specifically, the method first obtains an initial estimate of the attenuation map using a physics-based reconstruction approach. This initial estimate is then segmented into different regions using a DL-based segmentation approach. The segmented regions are assigned pre-defined attenuation coefficient values, yielding an attenuation map used to reconstruct MPS images. To evaluate this method, we conducted an IRB-approved retrospective study in patients administered MPS scans. We collected N = 648 anonymized clinical SPECT/CT stress MPS studies with SPECT projection data and CT images along with radiology reports. The dataset was divided into N = 508 training patients and N = 140 test patients. The segmentation network was trained using five-fold cross validation using the training set. For the evaluation study, we introduced synthetic cardiac defects of different extents, severities, and locations into the projection data of half of the patients in the test dataset using a defect-insertion approach (9). The SLAC method was evaluated on the task of defect detection with a multi-template model observer (10) with rotationally symmetric frequency channels (11,12). The performance was quantified using the area under the ROC curve (AUC) and was compared with the CTAC method and another method where no AC was performed (NAC). The workflow of the evaluation study is shown in Fig. 1a. We followed the best practices for evaluating AI algorithms (RELAINCE guidelines) to maintain the highest level of rigor in this study (13).

Results: The SLAC method was statistically non-inferior to the CTAC method on the defect detection task with the AUC difference within a margin of 3% of AUC (Fig. 1b). The ROC plots obtained with the SLAC and CTAC methods were approximately overlapping (Fig. 1c). Further, the SLAC method outperformed the NAC method. Further, in our sub-group analysis, the SLAC method yielded similar AUC values as the CTAC method for different defect extents and severities (Fig. 1d).

Conclusions: A scatter projection and DL-based tx-less AC method was observed to be non-inferior to a standard CTAC method on the task of detecting perfusion defects as evaluated using a retrospective study with anonymized clinical SPECT data of patients administered MPS studies with a model observer. The results motivate further clinical evaluation of the SLAC method.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint