Improved Deep Learning Model for Bone Age Assessment using Triplet Ranking Loss

Byeong-Uk Bae, Woong Bae, Kyu-Hwan Jung

Abstract: We propose a deep learning model using triplet ranking loss for bone age assessment. We build a hand segmentation network and transformation network for preprocessing to normalize x-ray images. We also added a triplet ranking loss to regression loss so that the embedded feature can be ordered. As the results, the model learns with ordered features show better performance. We evaluated our model with RSNA bone age assessment competition dataset.

Author Affiliation: Vuno Inc.

Keywords: bone age assessment, regression loss, triplet ranking loss, embedded feature