Development and evaluation of a radiomics model of resting 13N-ammonia positron emission tomography myocardial perfusion imaging to predict coronary artery stenosis in patients with suspected coronary heart disease

Figures

• 

  Figure 1. Flow chart of the study population. “n=112/n=101/n=9/n=3/n=2/n=42” is the numbers of patients and “n=126” is the number of coronary segments. CHD, coronary heart disease; MPI, myocardial perfusion imaging; CAG, coronary angiography.

• 

  Figure 2. Parameter analysis of MPI PET. A 59-year-old man presented with chest tightness for more than 1 month. Recent CAG suggested that the LCX had 30% stenosis and RCA occlusion. (A) The time/activity curves display radiotracer activity within 3 segments. (B) The polar map that is segmented into 3 shows rest MBF in mL/g/min. The MBFs of the LAD/LCX/RCA segment were 0.75/0.81/0.53 mL/g/min, respectively. (C,D) Hypoperfused myocardium is displayed in polar map coordinates (C) and 3D parametric surfaces (D) with 3 segments. TPD represents the EXT (black part) and the severity of the hypoperfused myocardium. The TPDs of the LAD/LCX/RCA segment were 0%/10%/63%, respectively. The EXT of the LAD/LCX/RCA segment were 2%/16%/78%, respectively. (E) The polar map was scored based on a 17-segment model and a categorical scale of 0–4 (0= normal; 4= absent perfusion). The SRSs for LAD/LCX/RCA were 0/1/11, respectively. MPI, myocardial perfusion imaging; PET, myocardial perfusion imaging; CAG, coronary angiography; LCX, left circumflex coronary artery; RCA, right coronary artery; MBF, myocardial blood flow; LAD, left anterior descending coronary artery; TPD, total perfusion deficit; EXT, defect…

• 

  Figure 3. Workflow of myocardial segmentation of MMI PET. (A) Load image. (B) Reposition the 3 positions of the heart (LV apex, LV base, and RV). (C) Search the ventricular boundaries and generate the VOI and MASK. (D) Three coronary segments were outlined on the LV polar map. (E) Generated VOI and MASK of the LAD/LCX/RCA segment. MMI, myocardial metabolic imaging; PET, myocardial perfusion imaging; LV, left ventricular; RV, right ventricular; VOI, volume of interest; LAD, left anterior descending coronary artery; LCX, left circumflex coronary artery; RCA, right coronary artery.

• 

  Figure 4. Heat map chart between radiomics features. Red indicates a positive correlation and blue indicates a negative correlation. The darker the color, the higher the correlation between the 2 features. (A) Cross-correlation of 996 features. (B) The cross-correlation of the remaining 211 features after Pearson correlation analysis.

• 

  Figure 5. Model selection and weights of features. (A) Rad model for feature selection; 12 radiomics features corresponding to the optimal validation AUC were selected. (B) Coefficients of 12 features in the Rad model. (C) MPI model for feature selection; 2 features that correspond to the optimal validation AUC were selected. (D) Coefficients of 22 features in the MPI model. Rad, radiomics; AUC, area under the curve; EXT, defect extent; MBF, myocardial blood flow; MPI, myocardial perfusion imaging.

• 

  Figure 6. Nomogram integrating the Rad and MPI models. The nomogram was developed from the data of the training set, which integrated the Rad and MPI scores. In the nomogram, we drew an upward vertical line according to the values of the Rad and MPI scores to determine the corresponding points. The sum of these 2 points was the total number of points. We drew a downward vertical line to determine the risk of coronary stenosis. Rad, radiomics; MPI, myocardial perfusion imaging.

• 

  Figure 7. Performance of the models in the training set and testing set. (A,B) ROC for predicting coronary stenosis. In the training (A) and testing (B) sets, the AUCs of the Rad model and nomogram were higher than that of the MPI model, but the AUC between the 2 was close. (C,D) Comparison of the AUC using the DeLong test. In the training (C) and testing (D) sets, the AUCs of the Rad model and nomogram were significantly different from that of the MPI model, but there was no significant difference between the AUC of the Rad model and nomogram. (E,F) Decision curve analysis. The y-axis measures the net benefit, which is calculated by adding true positives and subtracting false positives. The closer the curve is to the upper right, the higher the net benefit. In the training (E) and testing (F) sets, the Rad model and nomogram have a closer net benefit, which is higher than that of the MPI model. AUC, area under the curve; RAD, Rad model; NOM, nomogram; MPI, myocardial perfusion imaging; ROC, receiver operating characteristic curve; Rad, radiomics.

• 

  Figure 8. ROC for predicting coronary stenosis in the subgroup analysis. (A) Myocardial segment. (B) Number of stenotic vessels. (C) Left main trunk stenosis. (D) Cardiac risk factors. (E) Total dose of 13N-NH3. (F) Unit dose of 13N-NH3. LAD, left anterior descending coronary artery; LCX, left circumflex coronary artery; RCA, right coronary artery; ROC, receiver operating characteristic curve.