Adaptive Multi-Scale Strategy for Deep Reinforcement Learning Based 3D Ear Fenestra Vestibuli Landmark Localization in U-HRCT Images

The fenestra vestibuli is one of the main incidence sites of some ear diseases such as otosclerosis and neurological deafness. Therefore, the automatic localization of the fenestra vestibuli has become a key task in the intelligent analysis of 3D ear CT images. Localization methods based on deep reinforcement learning (DRL) usually adopts the fixed-scale multi-stage localization strategy, which is time-consuming due to training multiple scale models, and cannot deal with tiny targets well in abnormal samples. In this paper, an adaptive multi-scale strategy is proposed for 3D ear fenestra vestibuli landmark localization, which automatically adjusts the receptive field and step size of agent by introducing an adaptive scale vector into DRL algorithms. Accurate and robust localization results can be achieved after a single stage of training. Experimental results on the self-built fenestra vestibuli localization dataset show that, the proposed adaptive multi-scale localization strategy is well compatible with the existing DRL algorithms. Especially on the abnormal test set, compared with the fixed multi-scale localization strategy, the successful detection rate of the proposed strategy increased by 9.29% within the precision range of 2.60 mm, while the training time have decreased by 30%.