Denoising of Dual Energy X-ray Absorptiometry Images and Vertebra Segmentation

Sammanfattning: Dual Energy X-ray Absorptiometry (DXA) is amedical imaging modality used to quantify bone mineral density and to detect fractures. It is widely used due to its cheap cost and low radiation dose, however it produces noisy images that can be difficult to interpret for a human expert or a machine. In this study, we investigate denoising on DXA lateral spine images and automatic vertebra segmentation in the resulting images. For denoising, we design adaptive filters to avoid the frequent apparition of edge artifacts (cross contamination), and validate our results with an observer experiment. Segmentation is performed using deep convolutional neural networks trained on manually segmented DXA images. Using few training images, we focus on depth of the network and on the amount of training data. At the best depth, we report a 94 % mean Dice on test images, with no post-processing. We also investigate the application of a network trained on one of our databases to the other (different resolution). We show that in some cases, cross contamination can degrade the segmentation results and that the use of our adaptive filters helps solving this problem. Our results reveal that even with little data and a short training, neural networks produce accurate segmentations. This suggests they could be used for fracture classification. However, the results should be validated on bigger databases with more fracture cases and other pathologies.