関節リウマチの画像診断におけるAIの進展

タイトル: Pre-trained convolutional neural network with transfer learning by artificial illustrated images classify power Doppler ultrasound images of rheumatoid arthritis joints

概要: ObjectivesTo study the classification performance of a pre-trained convolutional neural network (CNN) with transfer learning by artificial images of the joint ultrasonography in rheumatoid arthritis (RA). MethodsWe focused on abnormal synovial vascularity and created 870 artificial ultrasound joint images based on the European League Against Rheumatism/Outcome Measure in Rheumatology scoring system. One CNN, the Visual Geometry Group (VGG)-16 was trained with transfer learning using the 870 artificial images for initial training and the original plus five additional images for second training. Actual joint ultrasound images obtained from patients with RA were used for testing our models. ResultsWe obtained 156 actual ultrasound joint images from 74 patients with RA. Our initial model showed moderate classification performance, but grade 1 was especially low (area under curve (AUC) 0.59). In our second model, grade 1 showed improvement (AUC 0.73). ConclusionsThis study was a novel attempt at using artificial joint images for training VGG-16. We concluded that artificial images were useful for training VGG-16. Use of artificial images might improve CNN training efficiency and allow development of the applications not only for ultrasound images but also other medical imaging modalities. KEY MESSAGESO_ST_ABSWHAT IS ALREADY KNOWN ON THIS TOPICC_ST_ABSThe application of artificial intelligence to rheumatology imaging has been actively studied. Previous studies have reported the application of convolutional neural network (CNN)-based scoring of synovial vascularity in ultrasound images of the hand of rheumatoid arthritis. The training process for medical artificial intelligence requires a large number of real-world clinical images, but collecting these images poses various challenges, including ethical concerns, and demands considerable effort. WHAT THIS STUDY ADDSIn our new approach, we used artificially drawn ultrasound images as training data for artificial intelligence. Using these images, we trained a CNN and investigated the models ability to classify actual clinical images. Our model demonstrated moderate classification performance on real clinical images and showed good agreement with human raters. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICYOur novel approach of using artificial images for training artificial intelligence has the potential to be applied in medical imaging fields that face difficulties in collecting real clinical images.