SRS94 - Using Residual Network (ResNet) convolutional neural networks to classify sarcopenia severity on two-dimensional magnetic resonance images of the foot in patients with diabetic foot disease

Abstract Introduction ResNet is a convolutional neural network (CNN) in the world of artificial intelligence and is used for image classification in the world of medicine. Initial ResNet structures (ResNet-18 and ResNet-34) had a basic CNN architecture whilst more recent models, ResNet-50 and ResNet-101 follow a bottleneck architecture. Diabetic foot disease (DFD) is a complex disease and is associated with lower limb amputation. Magnetic resonance imaging (MRI) is commonly used in patients with DFD. Aims To explore the use of deep learning models to identify predictors on MRI for limb loss in patients with DFD by using sarcoepenia as a potential surrogate marker. Methods Two-dimensional images of the foot at the base of the 1st metatarsal were classified as having mild, moderate or severe sarcopenia. A subset of 50 images were also graded by a musculoskeletal radiologist to establish the inter-rater reliability. 824 images were annotated. Following data pre-processing and data augmentation, 1740 images were available for the deep learning models which were split into a 70:20:10 ratio for training:validation:testing. Results The inter-rater reliability was 0.827 [95% c.i. 0.726–0.928; P-value < 0.001]. ResNet-18 had a 75% accuracy in classifying sarcopenia severity, ResNet-34 had a 78.7% accuracy. ResNet-50 had an accuracy of 79% with an F1 score of 74% and ResNet-100 had an accuracy of 80.7% with an F1 score of 69.3% in classifying sarcopenia severity on two-dimensional MRI pictures. Conclusions ResNet models are useful in classifying the severity of sarcopenia in patients with diabetic foot disease.