Keywords:
Bones, Paediatric, CAD, Digital radiography, Diagnostic procedure, Computer Applications-Detection, diagnosis, Metabolic disorders
Authors:
J. Song1, X. Fang1, Z. Yin2, Z. Xing2, P. Gong2, X. Li2, Y. Yu2, C. Gao1; 1Wuxi/CN, 2Beijing/CN
DOI:
10.26044/ecr2019/C-2335
Methods and materials
The left-hand radiographs of 5000 children with suspected metabolic disorders such as pubertas praecox were acquired from Wuxi Children’s Hospital(Fig.1).
From the total cases above,
1224 were randomly selected as training set (2~16years old: male 406,
female 818).
Another 101 patients were acquired prospectively and were continuously used as a validation set (2~16 years old: male 27,
female 74).
The scoring assignment method for determining bone age involves detailed analyses of the degree of growth of key epiphysis(Fig.2). So,
in the study several professional pediatric radiologists annotated the development stage of every key epiphysis in left-hand radiographs according to China 05. In order to reduce the influence of inconsistent opinions of different experts on training and testing the AI model,
two experts annotated training data with double-blind,
then we took the mean developmental stage of the experts’ results as the gold standard to train our AI model; In the same way ,
the validation data be annotated by 4 experts,
taking the mean developmental stage as the gold standard to validating our AI model.
After collecting and labelling the left-hand radiographs,
we adopted the method based on deep learning to automatically determine bone age.
Our automatic system followed the procedure of the China 05 method.
Specifically,
the first step,
we trained a cascade pyramid network which was used for the task of pose estimation in computer vision field to detect the key-points of epiphysis of 13 long or short bones .
After training,
the model could find all the key-points; so,
we successfully extracted the epiphyseal region of interests of the left-hand radiogram.
Next step,
we needed to train a regression model to evaluate the development stage of the key epiphysis.
The ground truth came from the label of the professional pediatric radiologist.
We adopted the ResNet50 as our backbone to extract the feature of the region of epiphysis,
then added 13 regression branches,
where each branch was used for regressing corresponding epiphysis’s development stage.
The results of the development stage predicted by the trained model of the validation set was that most of the predicted development stages approximated the labels of the professional pediatric radiologist.
Finally,we got the development stage of the epiphysis of 13 long or short bones,
used the China 05 method,
where each development stage of males and females was respectively replaced by a score. The total score was calculated to transform to the bone age of males and females,
to get the bone age of males and females respectively.(Fig.3)