Deep learning-based fully automatic segmentation of wrist cartilage in MR images

Abstract

The study objective was to investigate the performance of a dedicated convolutional neural network (CNN) optimized for wrist cartilage segmentation from 2D MR images. CNN utilized a planar architecture and patch-based (PB) training approach that ensured optimal performance in the presence of a limited amount of training data. The CNN was trained and validated in 20 multi-slice MRI datasets acquired with two different coils in 11 subjects (healthy volunteers and patients). The validation included a comparison with the alternative state-of-the-art CNN methods for the segmentation of joints from MR images and the ground-truth manual segmentation. When trained on the limited training data, the CNN outperformed significantly image-based and PB-U-Net networks. Our PB-CNN also demonstrated a good agreement with manual segmentation (Sørensen-Dice similarity coefficient [DSC] = 0.81) in the representative (central coronal) slices with a large amount of cartilage tissue. Reduced performance of the network for slices with a very limited amount of cartilage tissue suggests the need for fully 3D convolutional networks to provide uniform performance across the joint. The study also assessed inter- and intra-observer variability of the manual wrist cartilage segmentation (DSC = 0.78-0.88 and 0.9, respectively). The proposed deep learning-based segmentation of the wrist cartilage from MRI could facilitate research of novel imaging markers of wrist osteoarthritis to characterize its progression and response to therapy.

Keywords: applications, cartilage, human study, methods and engineering, musculoskeletal, postacquisition processing, quantitation.

FIGURE 1

Illustration of the manual segmentation results. (A) Preliminary delineation of the wrist joint area. (B) Final binary mask obtained after thresholding and manual correction

FIGURE 2

Schematic representation of data splitting for the different stages of the CNN development for the “hold-out” training/testing approach. DICOM, Digital Imaging and Communications in Medicine

FIGURE 3

Configuration of PB-CNN optimized for wrist cartilage segmentation

FIGURE 4

Dependence of DSC value on the training data amount and sample selection (“hold-out” training/testing approach). Blue dots correspond to consecutive inclusion of the data of each subject (from #1 to #5) to the training dataset (TD). Orange dots correspond to a random selection of slices for training in the indicated proportions from the whole TD (full TD). PB-U-Net, patch-based U-Net

FIGURE 5

Illustrations of performance of the proposed PB-CNN (red: correctly segmented pixels [true positives]; green: pixels incorrectly assigned to the background [false negatives]; and blue: pixels incorrectly assigned to the cartilage [false positives]). (A) Representative segmentation example (healthy subject, medial slice, DSC = 0.86, visual evaluation score = 8); zoomed-in cartilage area is shown in (B). The green arrow points to a vessel that had contrast and geometry similar to cartilage but was not assigned to this type of tissue by our PB-CNN. (C) Additional segmentation example (healthy subject, medial slice, DSC = 0.81, visual evaluation score = 6). (D) Example of segmentation of patient data with diminished performance (medial slice, DSC = 0.69, visual evaluation score = 3). The yellow circles show the skin tissue considered by CNN as cartilage. (E, F) Additional illustrations of CNN performance on the images of patients. The yellow arrows point to the high signal intensity lesions, which were correctly excluded by the proposed PB-CNN from the segmented mask