This study explored the feasibility and efficacy of radiomics methods when applied to LV tomograms from D-SPECT MPI for auxiliary diagnosis. At this stage, the experimental results show that the proposed method has a good ability to discriminate whether there is an ischemic area on LV tomograms with these specified conditions of image quality, training intensity, and sample number. The sensitivity and specificity exceeded 85%, the AUCs were all over 0.97, the agreement rates were above 90% and the NPVs and PPVs were above 87%, indicating that this method had above-average performance. When identifying the area of myocardial ischemia, three indexes, the Dice coefficient, IOU and Hausdorff distance, were also similar to those of the gold standard in this study, indicating that this method also achieves a desirable performance for identifying the morphology of ischemic areas. These results also illustrate that the radiomics method is feasible in assisting with the diagnosis of myocardial ischemia using LV tomograms obtained from D-SPECT MPI, and even though this method currently fails to achieve complete accuracy due to the limitations of the study and existing technology, this idea still has a promising future.
When diagnosing myocardial ischemia using nuclear medicine, LV tomograms are always an important diagnostic tool. Although QPS_QGS software can quantitatively analyze myocardial images, LV tomograms remain an important reference for nuclear medicine imaging radiologists diagnosing diseases. Previous radiomics studies using quantitative indexes and semiquantitative data of MPI (e.g., polar maps) have achieved some promising results. For example, a study conducted by Takayuki Shibutani et al. aimed to detect the accuracy of artificial neural networks in detecting areas with myocardial perfusion abnormalities on SPECT, but the method was mainly applied to polar maps . Three studies by Julian Betancur et al. used polar maps and quantitative data provided by QPS_QGS software: one study compared the ability of total perfusion deficit (TPD) and a DL model combining raw and quantitative perfusion polar maps to predict obstructive CAD ; one study used machine learning (ML) algorithms combining all relevant clinical variables and evaluated the benefit of predicting major adverse cardiac events (MACE) and the efficiency and accuracy of the algorithms for making diagnoses ; and the other study combined DL and TPD and assessed the ability of combined semiupright and supine stress MPI analysis to predict obstructive disease . Arsanjani et al.  used ML to integrate clinical data with quantitative features obtained from SPECT MPI images, but LV tomograms were only used for visual assessment. In this study, the possibility of applying radiomics for the diagnosis of CVD using nuclear medicine was studied from a new perspective, which has certain novelty.
The ResNet18 model chosen in this study falls under the category of DL, which is a branch of ML and is a part of the realization means of radiomics [23, 24]. This model, which has been recently developed, has resulted in more accurate and more stringent criteria than earlier congeneric models. This model also had a higher degree of automated processing. The methodology of radiomics is generally summarized as follows: image acquisition; segmentation of a region of interest (ROI); feature extraction from the ROI; and training and validation of the artificial intelligence (AI) model . Following the process described above, the process of “feature extraction - training and validation” was validated using this model, which greatly simplifies utilization of the model, thus not only increasing the radiologists’ acceptance of the model but also reducing the difficulty of system deployment.
The purpose of this study was to explore the possibility of using a radiomics method with LV tomograms obtained from D-SPECT MPI to assist in the diagnosis of myocardial ischemia in CAD and to provide new ideas and relevant experience in applying radiomics to diagnose CVD using nuclear medicine as well as developing related auxiliary diagnosis software so that clinicians can diagnose patients more quickly and accurately, thereby benefiting patients.
There are also some limitations in this study. First, lesion delineation necessitates a high reliance on manual processing because of the lack of available commercialized or open-source automated/semiautomated tools. Second, to define the edge of the ischemic area, only the most obvious boundary can be selected under the existing conditions; this will not only cause some cases in which ischemia is present but is not apparent in the images to be missed but may also cause both overestimation and underestimation of the ischemic extent. Third, limited by the performance of the chosen tool, polygonal shapes were used to simulate the margins of the ischemic area, but this method does not fully or accurately reflect the true morphology of ischemic areas. To reduce error, consultations with multiple nuclear medicine radiologists and quality control measures were adopted in this study, but the associated errors still cannot be avoided.