Open AccessEnd-to-end bone age assessment based on Attentional Region Localization
Author(s) -
Chunde Yang,
Liu Qing-shui,
Zhangyong Li,
Wei Wang,
Xinwei Li
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1771/1/012007
Subject(s) - computer science , artificial intelligence , segmentation , exploit , end to end principle , noise (video) , pattern recognition (psychology) , deep learning , annotation , classification of discontinuities , interference (communication) , computer vision , image (mathematics) , telecommunications , mathematics , mathematical analysis , channel (broadcasting) , computer security
Pediatric bone age assessment (BAA) is an extremely important clinical method to investigate endocrinology, genetic and growth disorders of adolescents. The current deep learning-based BAA scheme generally feeds the images into the training model, but ignores the local details of the skeleton images and does not exclude the noise around the images. Many methods train segmentation or detection networks to exploit local information, but requires a lot of manual annotation and additional costs. In this paper, we proposed an attentional region localization method for BAA to automatically localize the hand region and local regions without any additional annotations. First, an attentional hand location module (AHLM) was used to obtain a clearer hand region, which eliminates the interference noise of the original image. Then an attentional region generation module (ARGM) was used to extract the local attentional regions with high discriminant features, which can help to optimize the entire network framework. We integrate the entire network into an end-to-end structure by jointly optimizing the network through a shared backbone and fully connected layers. The effectiveness of the proposed attentional region localization method was evaluated on an open dataset Radiological Society of North America (RSNA) with an average absolute error (MAE) of 6.14, which performs better than most existing methods.