PROJECT TITLE :

Attention-Guided Discriminative Region Localization and Label Distribution Learning for Bone Age Assessment

ABSTRACT:

Bone age assessment, also known as BAA, is an extremely useful tool in the clinical setting because it can be utilized to diagnose endocrine and metabolic disorders that occur during the course of child development. Methods based on deep learning that are currently available for classifying bone age either take the entire image as their input or make use of local information by annotating additional bounding boxes or key points. However, training with the global image underutilizes information that is only available locally, and providing additional annotations is both expensive and subjective. In this paper, we present a proposal for an attention-guided approach to automatically localize the discriminative regions for BAA without the use of any additional annotations. To be more specific, we start by training a classification model to learn the attention maps of the discriminative regions. After that, we locate the hand region, the region with the highest level of discrimination (the carpal bones), and the region with the next highest level of discrimination (the metacarpal bones). Next, using the attention maps as a guide, we extract from the original image the parts of the image that contain the most relevant information and aggregate the results for BAA. We propose using joint age distribution learning and expectation regression instead of taking BAA as a general regression task, which is suboptimal due to the label ambiguity problem in the age label space. This method makes use of the ordinal relationship among hand images with different individual ages and results in more accurate age estimation. Taking BAA as a general regression task is suboptimal due to the label ambiguity problem in the age label space. On the RSNA pediatric bone age data set, a significant amount of testing is carried out currently. Our method achieves competitive results when compared to existing state-of-the-art deep learning-based methods that require additional manual annotations. This is possible because our method does not use any additional manual annotations.