Automated Bone Density Measurements using Deep Learning

Folle L, Meinderink T, Simon D, Liphardt AM, Thies M, Krönke G, Schett G, Kleyer A, Maier A (2022)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2022

Series: International Workshop on quantitative musculoskeletal imaging (QMSKI)

Book Volume: 23

Pages Range: O55

Conference Proceedings Title: 23

Event location: Nordwijk

Abstract

Introduction
Hand bone loss is characteristic of rheumatoid or psoriatic arthritis, which clinically manifests with pain, swelling, and stiffness of the affected joints. Typically, hand joints are affected already in the early stages of arthritis¹. High-resolution peripheral quantitative computed tomography (HR-pQCT) allows to visualize structural changes in the bone in 3D and enables measurements derived from the scans of the patient’s hand such as the volume of a bone. For this, a segmentation of the scans is necessary, which is reader-dependent and time-consuming, especially in challenging cases². This work aims to replace this manual expert segmentation with a neural network (figure 1).

Materials and methods

The segmentation network was trained using 541 HR-pQCT patient scans (RA, PsA, and OA patients with motion grades up to 3) as input and the corresponding expert segmentation of the second metacarpal bone as reference. To optimize the parameters of the network, the current prediction was compared with the expert segmentation using the Dice coefficient. Pre-processing steps included resizing, flipping of left hands, and intensity normalization. The architecture of the segmentation network is based on the 2D U-Net, which was pre-trained on ImageNet. The predictions were resized to the original resolution and converted to the scanner-specific data format as post-processing steps. The scanner then calculated the average bone-mineral density (D100) based on the scan and the corresponding network-generated segmentation. The scans were split into training, validation, and testing with ratios of 70%, 20%, and 10% respectively.

Results and discussion

The bone segmentation using the neural network on the test set reached a Dice coefficient of 97.2%. A quantitative comparison of the network prediction (blue) and the expert annotation (red) is presented in figure 2. The deviation of the expert annotation from the bone contour is likely due to the shift of the scan when exceeding the specified maximum number of slices of the scanner. Since the neural network operates on a slice level, such deviations are implicitly accounted for.

     The results of the vBMD analysis are summarized in figure 3. Significant Pearson correlation with 0.999 (p<.001) was achieved for the D100 on the test set. Still, a slight offset is present in the predictions, which might lead to underestimating the D100. However, this could be eliminated in a calibration step beforehand.

Conclusion

Neural networks can be used to generate accurate segmentations of the MCP2 head even with erosions present and can be integrated into the clinical workflow with a performance on par with expert annotators and save time through the automation of the annotation process.

Authors with CRIS profile

How to cite

APA:

Folle, L., Meinderink, T., Simon, D., Liphardt, A.-M., Thies, M., Krönke, G.,... Maier, A. (2022, June). Automated Bone Density Measurements using Deep Learning. Poster presentation at QMSKI, Nordwijk, NL.

MLA:

Folle, Lukas, et al. "Automated Bone Density Measurements using Deep Learning." Presented at QMSKI, Nordwijk Ed. Bert van Rietbergen, Harry van Lenthe, Quentin Grimal, 2022.

BibTeX: Download