SVTNet : Automatic bone age assessment network based on TW3 method and vision transformer

Abstract

Abstract This study aims to develop a proficient and clinically applicable algorithm that can accurately assess bone age. This algorithm is based on the principles of the Tanner‐Whitehouse 3 (TW3) integral approach, and aims to achieve efficiency, scalability, and interpretability. We developed a model for bone age prediction in children. The model was tested on a pediatric dataset from a tertiary care hospital consisting of left‐hand radiographs of children between the age of 0 and 18. Our model consists of removing the arm portion using a pre‐trained YOLO network, localizing 37 key points in the hand bone portion using a spatial configuration network, and segmenting the original image through 20 of these points to obtain 20 fixed‐size patches. Finally, each of the 20 bone images is classified by training a visual transformer (ViT) model. In this study, a hybrid network, SVTNet, was developed that incorporates visual transformers to obtain estimates of bone age in the carpal (C series) and metacarpal (RUS series) bones. The sum of the clinical TW3 scoring region scores and bone maturity scores were utilized to determine the bone age for each corresponding region. The performance of the algorithm was evaluated in terms of both training and testing by evaluating 3871 left hand X‐ray micrographs obtained from a tertiary hospital in China. The results showed that the average absolute error of bone age estimation was 0.50 years for the RUS series of bones and 0.47 years for the C series of bones. The main contribution of this study is to propose, for the first time, a ViT‐based bone age assessment method that automates the entire process of the TW3 algorithm and is clinically interpretable, with predictive accuracy comparable to that of an experienced orthopedic surgeon.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 13 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.