Improved The Diagnostic Accuracy Of Stress SPECT-myocardial Perfusion Imaging For Obstructive Coronary Heart Disease Based On Transfer Learning

Objective:Three pre-trained models,ResNet34,ResNet50 and ResNet101,were obtained based on Transfer learning,and trained using stress myocardial perfusion imaging（MPI）perfusion polar maps combined with patient gender characteristics.The best of the three pre-trained models was compared with the standard quantitative parameters of Total Perfusion Defect（TPD）and physician’s subjective visual assessment to investigate the effect of different number of convolution layers on the diagnostic performance of the models.The analysis was performed to verify the value of migration learning in MPI researchMethods:1.A single-center retrospective analysis of 326 patients who underwent stress SPECT-MPI for suspected CAD at the Department of Nuclear Medicine,First Hospital of Shanxi Medical University from March 2020 to March 2023 was selected,including226 male patients and 100 female patients.Clinical information such as age,gender,body mass index（BMI）,history of smoking,Hypertension,Hyperlipidemia and Diabetes Mellitus and imaging data were also collected,and the patients were divided into a positive obstructive CAD group（n=183）and a negative obstructive CAD group（n=143）according to the results of Coronary Angiography.2.TPD is a quantitative indicator of the extent and severity of myocardial perfusion defects and is automatically calculated by the Cedars quantitative perfusion SPECT（QPS）software;the subjective visual assessment was performed independently by two nuclear medicine physicians,including a senior physician,visual semi-quantitative analysis was performed according to the 17-segment left ventricular myocardial model recommended by the American Board of Clinical Cardiology Committee on Cardiac Imaging,during which both physicians had access to all clinical information and imaging data except for CAG results.When two doctors disagree on a diagnosis,one of the senior doctors makes the final decision.3.The dataset was cropped by a nuclear medicine physician prior to training using Adobe Photoshop software without compressing the image quality,and the resolution was set to 224×224 pixels for MPI perfusion polar maps.The dataset was then randomly divided into training（n=194）,validation（n=66）,and test（n=66）sets according to the ratio of 6:2:2.In addition,to prevent overfitting of the ResNet model during training,the perfusion polar maps of each patient in the training set were augmented with data using scaling,horizontal flipping and vertical flipping methods.4.The learning process of the ResNet model with three different convolutional layers is divided into three steps:firstly,features are extracted from MPI perfusion target heart maps using convolutional layers with the number of iterations set to 30 and the learning rate set to 0.0001.Secondly,patient gender features are added to the fully connected layer to distinguish the image differences between males and females under non-attenuation corrected conditions.Three parallel dichotomous outputs were provided after the fully connected layer to judge whether each of the three coronary vessels reached the degree of stenosis for diagnosing obstructive CAD.Finally,the dichotomous results for the three vessels are used to complete the patient-level prediction by or arithmetic.5.The three ResNet models were compared by using the differences of Area under Curve（AUC）of Receiver Operating Characteristic（ROC）and combined with precision,accuracy,sensitivity,specificity,recall and F1 score.The robustness of the model is evaluated by calculating the changes of training accuracy and training loss during the iterative process.The best ResNet prediction model is further compared and analyzed with the standard quantitative parameter stress TPD and physician subjective visual assessment.Results:1.There were no statistically significant differences between patients in the positive and negative obstructive CAD groups in terms of age,BMI,hypertension,hyperlipidemia,and history of diabetes mellitus（P>0.05）;whereas statistically significant differences existed in terms of gender and smoking history（P<0.05）.There were no statistical differences in clinical characteristics and coronary angiographic characteristics among the three groups of patients in the training,validation and test sets（P>0.05）.2.After 30 iterations,the ResNet101 model obtained the highest training accuracy（86.35%）and the lowest training loss（20.58%）,with better robustness compared to the other two ResNet models.In predicting obstructive CAD,the ResNet101 model significantly outperformed the remaining two ResNet models with fewer convolutional layers at the individual patient level and for LAD lesions as the number of convolutional layers increased(AUC _patients:0.911 vs 0.888 vs 0.871;AUC_LAD:0.905 vs 0.893 vs 0.851)However,for the diagnosis of LCX lesions,the predictive performance of all three ResNet models was poor(AUC_LCX:0.658 vs 0.585 vs 0.547);the predictive performance of the ResNet101 model for RCA lesions was higher than that of the ResNet34 model,but there was no significant improvement over the predictive power of the ResNet50model(AUC_RCA:0.802 vs 0.784 vs 0.773),and the diagnostic accuracy（84.85%vs87.88%vs 83.33%）and precision（70.59%vs 90.91%vs 68.75%）were significantly lower than those of the ResNet50 model.3.In the subsequent comparative analysis,we found that the ResNet101 model had the same diagnostic sensitivity of 89.19%as the standard quantitative parameter stress TPD for obstructive CAD,but in terms of overall predictive performance,the ResNet101model（AUC=0.911）was better than the stress TPD（AUC=0.894）and the physician’s subjective visual assessment（AUC=0.881）.Conclusion:This study employs transfer learning to investigate the differences in training robustness and predictive performance of ResNet34,ResNet50,and ResNet101 models with distinct convolution layers.Through comparative analysis using standard quantitative parameters,such as True Positive Detection（TPD）,and subjective visual evaluation by physicians,the selected ResNet101 model demonstrated superior diagnostic performance in predicting obstructive CAD.These findings highlight the significance of evaluating the impact of network depths on prediction outcomes before constructing deep learning（DL）models for specific classification tasks.Hence,selecting the most suitable model architecture can enhance the accuracy of intelligent diagnosis.This study underscores the importance of selecting an appropriate network architecture in the development of DL models for medical image classification tasks.