AIMS: To develop a deep learning model for classifying frames with vs. without optical coherence tomography (OCT)-derived thin-cap fibroatheroma (TCFA). METHODS AND RESULTS:Total 602 coronary lesions from 602 angina patients were randomized into training and test sets at a 4:1 ratio. A DenseNet model was developed to classify OCT frames with or without OCT-derived TCFA. Gradient-weighted class activation mapping was used to visualize the area of attention. In the training sample (35,678 frames of 480 lesions), the model with 5-fold cross-validation had an overall accuracy of 91.6±1.7%, sensitivity of 88.7±3.4%, and specificity of 91.8±2.0% (averaged AUC=0.96±0.01) in predicting the presence of TCFA. In the test samples (9,722 frames of 122 lesions), the overall accuracy at the frame level was 92.8% within the lesion (AUC=0.96) and 91.3% in the entire OCT pullback. The correlation between the %TCFA burdens per vessel predicted by the model compared with that identified by experts was significant (r=0.87, p<0.001). The region of attention was localized at the site of the thin cap in 93.4% of TCFA-containing frames. Total computational time per a pullback was 2.1 ± 0.3 seconds. CONCLUSIONS: Deep learning algorithm can accurately detect an OCT-TCFA with a high reproducibility. The time-saving computerized process may assist clinicians to easily recognize high-risk lesions and to make decisions in the catheterization laboratory.
RCT Entities:
AIMS: To develop a deep learning model for classifying frames with vs. without optical coherence tomography (OCT)-derived thin-cap fibroatheroma (TCFA). METHODS AND RESULTS: Total 602 coronary lesions from 602 anginapatients were randomized into training and test sets at a 4:1 ratio. A DenseNet model was developed to classify OCT frames with or without OCT-derived TCFA. Gradient-weighted class activation mapping was used to visualize the area of attention. In the training sample (35,678 frames of 480 lesions), the model with 5-fold cross-validation had an overall accuracy of 91.6±1.7%, sensitivity of 88.7±3.4%, and specificity of 91.8±2.0% (averaged AUC=0.96±0.01) in predicting the presence of TCFA. In the test samples (9,722 frames of 122 lesions), the overall accuracy at the frame level was 92.8% within the lesion (AUC=0.96) and 91.3% in the entire OCT pullback. The correlation between the %TCFA burdens per vessel predicted by the model compared with that identified by experts was significant (r=0.87, p<0.001). The region of attention was localized at the site of the thin cap in 93.4% of TCFA-containing frames. Total computational time per a pullback was 2.1 ± 0.3 seconds. CONCLUSIONS: Deep learning algorithm can accurately detect an OCT-TCFA with a high reproducibility. The time-saving computerized process may assist clinicians to easily recognize high-risk lesions and to make decisions in the catheterization laboratory.