1. Deep learning-based localization and segmentation of wrist fractures on X-ray radiographs.
- Author
-
Joshi, Deepa, Singh, Thipendra P., and Joshi, Anil Kumar
- Subjects
- *
WRIST fractures , *ARTIFICIAL neural networks , *WRIST , *WRIST joint , *RADIOGRAPHS , *CARPAL bones - Abstract
X-rays are the primary tools in examining the suspected fractures in humans in this era. Manual examination of X rays is a time-consuming process, which requires expert radiologists or trained orthopaedic surgeons. The demanding workloads, unavailability of radiologists in small setups and at primary/community health centres results in the inability to diagnose and delay in the treatment due to unnecessary referrals by the primary care clinicians. Moreover, the shortage of expert radiologists and orthopaedic surgeons in medically under-resourced areas such as rural areas in India have motivated us to develop an automated fracture detection model. We have developed a deep neural network to detect, localize and divide the wrist region into segments to identify fractures around wrist joint in radiographs. The orthopaedic surgeon has manually annotated the fractures by drawing a bounding box and segmented mask. We have utilized two different domains of datasets for the better convergence of the model. There is a similarity in the crack patterns of the surface crack image dataset and the wrist fracture dataset, which consists of 3000 and 315 images. A part of the dataset is made publicly available for research purposes to overcome data collection and labelling barriers for identifying wrist fractures. The crack patterns in the wrist bone samples are learned by effectively transferring the knowledge or weights obtained from surface crack datasets. The proposed architecture utilizes Feature Pyramid Network as the backbone architecture where the last-level max pool layer of the architecture is replaced with the concatenation of AdaptiveConcatPool, AdaptiveAvgPool, AdaptiveMaxPool layers. Additionally, the concept of freezing and unfreezing the network during two phases of transfer learning is utilized for better model convergence. Every radiograph is assigned a ground truth label for evaluating the accuracy of the model. The performance measure for fracture detection and localization is evaluated using the Average precision value using the concept of Intersection over Union. The result of the proposed model is compared against the ground truth label annotated by the radiologists and related studies. The average precision of 92.278 on 50° and 79.003 on a strict scale of 75° was reported for fracture detection. Similarly, the average precision of 77.445 on 50° and 52.156 on a strict scale of 75° was reported for fracture segmentation. [ABSTRACT FROM AUTHOR]
- Published
- 2022
- Full Text
- View/download PDF