Deeplasia: prior-free deep learning for pediatric bone age assessment robust to skeletal dysplasias

Abstract

Abstract Background Skeletal dysplasias collectively affect a large number of patients worldwide. The majority of these disorders cause growth anomalies. Hence, assessing skeletal maturity via determining the bone age (BA) is one of the most valuable tools for their diagnoses. Moreover, consecutive BA assessments are crucial for monitoring the pediatric growth of patients with such disorders, especially for timing hormone treatments or orthopedic interventions. However, manual BA assessment is time-consuming and suffers from high intra-and inter-rater variability. This is further exacerbated by genetic disorders causing severe skeletal malformations. While numerous approaches to automatize BA assessment were proposed, few were validated for BA assessment on children with abnormal development. Objective We design and present Deeplasia, an open-source prior-free deep-learning approach for pediatric bone age assessment specifically validated on patients with skeletal dysplasias. Materials and methods We extensively experiment with training multiple convolutional neural network models under various conditions and select three to build a precise model ensemble. We utilize the public RSNA BA dataset consisting of training, validation, and test subsets each containing 12,611, 1,425, and 200 hand X-rays, respectively. For testing the performance of our model ensemble on dysplastic hands, we retrospectively collected 568 X-ray images from 189 patients with molecularly confirmed diagnoses of seven different genetic bone disorders including Achondroplasia and Hypochondroplasia. Results On the public RSNA test set, we achieve state-of-the-art performance with a mean absolute difference (MAD) of 3.87 months based on the average of six different reference ratings. We demonstrate the generalizability of Deeplasia to the dysplastic hands (unseen by the models) achieving a MAD of 5.84 months w.r.t. to the average of two reference ratings. Further, using longitudinal data from a subset of the dysplastic cohort (149 images), we estimate the test-retest precision of our model ensemble to be at least at the human expert level (2.74 months). Conclusion We conclude that Deeplasia suits assessing and monitoring the BA in patients with skeletal dysplasia.

Ähnliche Arbeiten