Ossification area localization in pediatric hand radiographs using deep neural networks for object detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Radiology

Background:

• Detecting ossification areas in hand X-rays is crucial for automated bone age estimation.
• Deep neural networks (DNNs) are effective but require large annotated datasets.
• Finetuning pre-trained DNNs offers a potential solution for data-scarce scenarios.

Purpose of the Study:

• To evaluate the effectiveness of finetuning a pre-trained DNN for detecting ossification areas in pediatric hand X-rays.
• To assess the impact of dataset size on the performance of finetuned DNNs.
• To compare the performance of finetuned DNNs against a network trained from scratch.

Main Methods:

• A Faster R-CNN network pre-trained on the COCO dataset was finetuned using 240 annotated pediatric hand radiographs.
• Subsampling techniques were employed to analyze the effect of varying training data sizes.
• Performance was evaluated using Intersection-over-Union (IoU), mean Average Precision (mAP@0.5IoU), and F1-Score against expert annotations.

Main Results:

• The finetuned network achieved a high mean Average Precision (mAP@0.5IoU) of 92.92 ± 1.93.
• Performance generally improved with more data, except for the wrist region.
• A network trained from scratch failed to produce accurate results, highlighting the benefit of pre-training.

Conclusions:

• Finetuning a pre-trained deep neural network with a small dataset (240 radiographs) is a successful strategy for detecting ossification areas in pediatric hand X-rays.
• The proposed method demonstrates high accuracy and robustness, comparable to expert radiologists.
• This approach offers a viable solution for automated bone age estimation in data-limited settings.