Chronological age estimation for medico-legal expertise-based on sternoclavicular joint CT images using a deep neural network

Abstract

The aim of this study was to develop and validate fully automated deep learning models to estimate chronological age from sternoclavicular CT images to help forensic age estimation and understand its limitations. A total of 742 whole-body CT and 164 pediatric chest-abdomen-pelvis CT scans (age: 1–60y, 437 m and 469f) were collected as a training dataset. A deep learning pipeline was implemented to segment the clavicle volume of interest, train an age estimation model, and finally fine-tune the network. The predictive performance of nine deep learning models was assessed and compared using 5-fold cross-validation. A transfer learning experiment was designed to evaluate the generalizability of the pre-trained models, using a fine-tuning group (age: 15–35y, 6 m and 4f) and a validation group (age: 16–35y, 6 m and 4f). Clinical age assessment based on clavicle bone was conducted on 5 thorax CT scans (4 m and 1f, age: 16–32y) and 5 sternoclavicular joint CT scans (unknown age) by one radiologist and two forensic pathologists. The intra- and inter-observer agreement of experts was assessed. A mean absolute error (MAE) of 4.23 ± 4.49 years, an area under the receiver operating characteristic (AUC) of 0.99 for age classification ($>14$ years and $>18$ years) and an accuracy of 0.97 for classification of ossification stages were achieved in the cross-validation. An MAE of 3.30 ± 3.58 years and an accuracy of 0.90 for ossification stage classification were achieved after fine-tuning. The three experts disagreed on the images that met the diagnostic requirements in 2 cases. Intra-observer agreement varied between experts. This study concluded that a fully automated deep neural network, employing a transfer learning strategy, exhibits potential for estimating chronological age from clavicular CT images.