An Automatic Assessment Model of Adenoid Hypertrophy in MRI Images Based on Deep Convolutional Neural Networks

Adenoid hypertrophy is a pathological condition characterized by the enlargement of the adenoids in children, which may lead to various problems. The conventional manual measurement is time-consuming and subject to subjective errors. Previous automated methods based on landmark detection in X-ray images have shown reliability. However, these methods neglect global features, making it difficult to accurately locate landmarks and thus affecting the estimation of the adenoid-to-nasopharyngeal (AN) ratio when migrating to MRI images. In this paper, we first apply a deep-learning method to the automatic assessment of adenoid hypertrophy in MRI images and propose an adenoid network (ADNet) to capture local and global features near landmarks to achieve accurate landmark localization. Specifically, ADNet uses an encoder-decoder architecture where we employ a depthwise separable convolution-based encoder to extract local features, and then employ an adaptive convolution-based decoder to capture global features. We collected a dataset of 500 cephalometric MRI images to train and evaluate the performance of the proposed model. Our experimental results demonstrate that our network achieves state-of-the-art performance with an average radial error of 4.85 pixels for landmark detection, an average successful detection rate of 96% within 15 pixels, and an error of 0.026 for the calculated AN ratio.