Improving the Accuracy of Deep Learning Networks for Bone-Age Estimation by Incorporating Radiological Insight Guided Feature Analysis

(RSNA 2018, Wed Nov 28 2018 9:00AM – 9:10AM ROOM Z09)

PURPOSE

The Greulich-Pyle (GP) method of bone age determination primarily involves estimation of ossification of the epiphyseal centers around the radiocarpal, carpometacarpal and the proximal interphalangeal joints and the carpal bones. This radiological insight was applied to devise a novel sequential approach, involving segmentation of relevant wrist anatomy from hand X-Rays followed by deep learning, and compared it against a standard deep learning technique to assess bone age in pediatric population between 7 years to 18 years.

METHOD AND MATERIALS

Dataset containing 12,600 radiographs provided by RSNA for Bone Age Challenge is used for this work. Out of the 12,600 we used ~10,000 radiographs of children between 7 years to 18 years of age. The intensity values of the hand radiographs are standardized across dataset by histogram matching image pre-processing techniques. A pre-processing algorithm was created to crop relevant regions, i.e. proximal phalanges, metacarpals, carpals and distal ends of radius and ulna, from the hand radiographs. Finally, ~9,000 cropped images were used to train a convolutional neural network implemented in the research version of HealthSuite Insights (Philips HealthTech) to predict the bone age from the image. The remaining images (~1,000) were used for validation purposes. Additional datasets of 200 test images released by RSNA and 50 test images obtained as part of routine clinical practice (extracted from PACS and anonymised using HIPAA compliant methods) were used to calculate Mean Absolute Error (MAE). Similar MAE was calculated for the same convolutional neural network implemented in the research version of HealthSuite Insights (Philips HealthTech) trained on the images without cropping the images.

RESULTS

The performance of deep learning model trained using cropped images was found to be superior with MAE of 5.08 months compared to the model trained using the full radiographs, which had MAE of 5.51 months.

CONCLUSION

The accuracy of machine learning models for specific tasks on radiographs can be improved by training using cropped/segmented radiographs containing areas of anatomy relevant to the task.

CLINICAL RELEVANCE/APPLICATION

Automated and accurate bone age estimation has wide-ranging clinical and medicolegal applications. Our novel method combines radiologist guided feature-based analysis with deep learning to improve the accuracy.