Electronic ISSN 2287-0237

VOLUME

PERFORMANCE OF CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING FOR SKELETAL BONE AGE ASSESSMENT

FEBRUARY 2019 - VOL.15 | ORIGINAL ARTICLE
OBJECTIVES:

Bone age assessment is used by clinicians for estimating the maturity of a child’s skeletal system. Traditionally, physicians use template matching methods (GP and/or TW2). Time and accuracy of the evaluation rely on a physician’s experience. Therefore, this research proposes a fully automatic system for bone age assessment with cutting edge artificial Intelligence (AI) technology. 

MATERIAL AND METHODS:

Convolutional Neural Network (CNNs), a Deep Learning (DL) technique is applied to skeletal bone age prediction combined with transfer learning algorithm. Hence, various kinds of transfer learning algorithms (ResNet-50, Inception-V3, and VGG-16) are investigated in training in the proposed model fed by a number of x-ray images (12,000 image approximately—imbalanced data). 

RESULT: VGG-16 shows significant accuracy compared to ResNet-50 and Inception-V3 (mae = 6.53, 20.52 and 43.11 months respectively) 

CONCLUSION:

The most effective pre-trained layer for CNNs in bone age assessment is VGG-16 according to the accuracy of its prediction. 

Keywords:

deep learning, convolutional neural network, bone age, growth disorder, maturity estimation, transfer learning

Received: January 11, 2019

Revision received: January 13, 2019  

Accepted after revision: February 11, 2019  

BKK Med J 2019;15(1): 1-6.

DOI: 10.31524/bkkmedj.2019.02.001

 

MEDIA
Figure 1: The proposed architecture of CNNs model for bone age assessment.
Figure 2: The X-Ray images distribution for bone age (0-240 months) of male (green) and female (blue).
Table 1: The 20 non-overlapped classes for the resampling.
Figure 3: Examples of image augmentation including rotation, scaling, and translation.
Figure 4: The proposed bone age assessment model with CNNs utilizing transfer learning, as well as, attention mechanism.
Table 2: The evaluation of different pretrained models
Figure 5: Example of the low quality images.
Figure 6: Evaluation of VGG-16.
Figure 7: Evaluation of ResNet-50.
Figure 8: Evaluation of Inception-V3.
SEARCH VOLUME & NUMBER
GO
RECENT VOLUME & NUMBER