mojtaba sirati-amsheh

Accurate bone age assessment is essential for determining the actual degree of development and indicating a disorder in growth. While clinical bone age assessment techniques are time-consuming and prone to inter/intra-observer variability, deep learning-based methods are used for automated bone age estimation. The current study aimed to develop an unsupervised pre-training approach for automatic bone age estimation, addressing the challenge of limited labeled data and unique features of radiographic images of hand bones. Bone age estimation is complex and usually requires more labeling data. On the other hand, there is no model trained with hand radiographic images, reused for bone age estimation. In this fundamental-applied research, the collection of Radiological Society of North America (RSNA) X-ray image collection is used to evaluate the efficiency of the proposed bone age estimation method. An autoencoder is trained to reconstruct the original hand radiography images. Then, a model based on the trained encoder produces the final estimation of bone age. Experimental results on the Radiological Society of North America (RSNA) X-ray image collection achieve a Mean Absolute Error (MAE) of 9.3 months, which is comparable to state-of-the-art methods.  This study presents an approach to estimating bone age on hand radiographs utilizing unsupervised pre-training with an autoencoder and also highlights the significance of autoencoders and unsupervised learning as efficient substitutes for conventional techniques.

Hand radiographs are commonly used to evaluate bone maturity. So that the significant difference between the estimated bone age and the chronological age can indicate a developmental disorder. However, the manual evaluation of images is usually a time-consuming and observer-dependent process. Therefore, in this paper, an automatic method for the assessment of bone age using radiographs of children's hands is proposed.

In this fundamental-applied research, the collection of radiographic images of the Radiological Society of North America (RSNA) was used, and transfer learning methods were proposed. The input images were first pre-processed due to low quality. Then a pre-trained model based on DenseNet-121 was used to extract the discriminating spatial features.

Evaluations using five pre-trained models on the RSNA dataset showed that the DenseNet-121 model, after adjustment, could perform better than other models, with a mean absolute error of 9.8 months.

Skeletal maturity can be estimated with satisfactory accuracy using the DenseNet-121 model, and this method can help radiologists in quick and accurate measurement of bone age.