Attention-Based Multiscale Localization for Bone Age Assessment

Abstract

Bone Age Assessment (BAA) is crucial for the biological maturity evaluation of children. Developing automated techniques of BAA has gained a lot of attention from both academia and medicine. This paper presents a novel deep-learning-based BAA method including refining and multi-scale processing of hand X-ray images. The refining step removes unnecessary background and noises in the images, resulting in a high-quality dataset. Such image refinement is beneficial for Region of Interest (ROI) localization step with self-attention mechanism. The localization model is trained with multiscale hand X-ray images separately to obtain multiple ROIs for each image. Eventually, the multiscale ROIs are used as complementary features of an image in training a regression model for BAA. We evaluate the performance of the proposed method using 2017 RSNA pediatric bone age challenge dataset. The results show mean absolute error of 3.52, which is 24.9% lower than state-of-the-art results.

Ähnliche Arbeiten