Auto-metric distribution propagation graph neural network with a meta-learning strategy for diagnosis of otosclerosis

Otosclerosis is a multifactorial bone disorder that affects the otic capsule; otosclerosis is a significant cause of deafness in adults. Since the lesion areas are frequently subtle, the diagnosis of otosclerosis on temporal bone CT images tends to be difficult, especially for fenestral otosclerosis. We design a deep learning model for diagnosing otosclerosis on CT scans in the case of limited samples. That is, we design a dual graph network, namely, ADP-GNN, for predicting otosclerosis-positive and otosclerosis-negative samples; the network consists of point graphs and distribution graphs. More specifically, the point graph is used to model the instance-level relation between nodes, and the risk factors are integrated into it for multimodal diagnosis. The distribution graph is used to model the distribution-level relation between samples, and the copula function is introduced to better measure the dependency between nodes. The autometric strategy is also used to make the model more flexible and to enable the sample to be evaluated independently. Through the propagation between the two graphs and metatraining, the labels of unknown nodes can be predicted. Test experiments on otosclerosis datasets show that the performance of our model achieves accuracies of 98.15