Diagnosis of Osteoporosis using modified U-net architecture with attention unit in DEXA and X-ray images

Osteoporosis, a silent killing disease of fracture risk, is normally determined based on the bone mineral density (BMD) and T-score values measured in bone. However, development of standard algorithms for accurate segmentation and BMD measurement from X-ray images is a challenge in the medical field.

The purpose of this work is to more accurately measure BMD from X-ray images, which can overcome the limitations of the current standard technique to measure BMD using Dual Energy X-ray Absorptiometry (DEXA) such as non-availability and inaccessibility of DEXA machines in developing countries. In addition, this work also attempts to analyze the DEXA scan images for better segmentation and measurement of BMD.

This work employs a modified U-Net with Attention unit for accurate segmentation of bone region from X-Ray and DEXA images. A linear regression model is developed to compute BMD and T-score. Based on the value of T-score, the images are then classified as normal, osteopenia or osteoporosis.

The proposed network is experimented with the two internally collected datasets namely, DEXSIT and XSITRAY, comprised of DEXA and X-ray images, respectively. The proposed method achieved an accuracy of 88% on both datasets. The Dice score on DEXSIT and XSITRAY is 0.94 and 0.92, respectively.

Our modified U-Net with attention unit achieves significantly higher results in terms of Dice score and classification accuracy. The computed BMD and T-score values of the proposed method are also compared with the respective clinical reports for validation. Hence, using the digitized X-Ray images can be used to detect osteoporosis efficiently and accurately.