A Comprehensive Survey on Bone Segmentation Techniques in Knee Osteoarthritis Research: From Conventional Methods to Deep Learning

Abstract

Knee osteoarthritis (KOA) is a degenerative joint disease, which significantly affects middle-aged and elderly people. The majority of KOA is primarily based on hyaline cartilage change, according to medical images. However, technical bottlenecks such as noise, artifacts, and modality pose enormous challenges for an objective and efficient early diagnosis. Therefore, the correct prediction of arthritis is an essential step for effective diagnosis and the prevention of acute arthritis, where early diagnosis and treatment can assist to reduce the progression of KOA. However, predicting the development of KOA is a difficult and urgent problem that, if addressed, could accelerate the development of disease-modifying drugs, in turn helping to avoid millions of total joint replacement procedures each year. In knee joint research and clinical practice there are segmentation approaches that play a significant role in KOA diagnosis and categorization. In this paper, we seek to give an in-depth understanding of a wide range of the most recent methodologies for knee articular bone segmentation; segmentation methods allow the estimation of articular cartilage loss rate, which is utilized in clinical practice for assessing the disease progression and morphological change, ranging from traditional techniques to deep learning (DL)-based techniques. Moreover, the purpose of this work is to give researchers a general review of the currently available methodologies in the area. Therefore, it will help researchers who want to conduct research in the field of KOA, as well as highlight deficiencies and potential considerations in application in clinical practice. Finally, we highlight the diagnostic value of deep learning for future computer-aided diagnostic applications to complete this review.

Keywords: bone segmentation; deep learning; knee osteoarthritis; machine learning; segmentation.

Figure 1

Methods of classifying the knee bones [12].

Figure 2

Knee bone segmentation has benefits over other tissues because of its location and anatomical size. (a) Illustrates the MR image of a knee joint—patella, femur, and tibia bones, readily apparent with the accompanying cartilage surfaces. TB = tibia bone, FB = femoral bone, PC = patellar cartilage, FC = femoral cartilage, TC= tibia cartilage. (b) Shows segmented tibia (TB) and femur (FB), which usually have better demarcation [19].

Figure 3

Illustrates the search with ASM for the face. (a) In the case of a point being near the target; (b) shows how the ASM can break down if the starting position is too far from the target [27].

Figure 4

An example illustrates both ROI detection failures/recoveries and leak detection and correction. (a) An MRI image; (b) the ROI block is detected with only the femur bone detected, not the tibia; (c) after lowering the ROI detection threshold, both bones are detected; (d) mask for the GC output; (e) after morphological processes; (f) the resulting two potential skeletons, with a leak seen in the tibia bone; (g) the tibia bone has a leak that connects fat and other tissues to the tibia; (h) initial step in detecting a leak is to use a morphological opening; (i) residual content resulting from subtracting (h) from (g). (j) Following an examination of the remains in (i), the leak detection method identifies a leak and decides that only the pixels in the leak are affected; (j) are relevant to the tibia (k) after adding the appropriate pixels in (j) to (h), resulting in a leak-free tibia (i), the femoral mask (l) and (m). After applying the morphological aperture to check for leakage (n) the remaining pixels after subtracting (m) from (i). On this basis, it is concluded that there is no leak, and the pixels are reinserted (o). (o) femur and tibia masks as a result (p) GC segmentation in white and manual segmentation in yellow determined with DICE = 0.95 and 0.96 resolution for femur and tibia bones, respectively [45].

Figure 5

One sample slice’s bone segmentation in coronal view. (a) Original image; (b) multi-atlas-based spatial prior; (c) segmentation result; (d) expert segmentation [49].

Figure 6

Analysis of the results of a general region-growing algorithm: (a) Original image (red arrow points to things (teeth and auris) to be segmented); (b) segmentation results from region growing using parameters r = 30; (c) results obtained using the robust split-and-merge algorithm; (d) the results showed that the edges of the images are more exact and smoother [53].

Figure 7

A typical system of machine learning [61].

Figure 8

Segmentation findings for MR images using the hybrid SVM-DRF model with five types of feature vectors [68].

Figure 9

Knee bone segmentation using (a) classical machine learning and (b) deep learning. Classic machine learning feature architecture includes hand-picked representations and mapping, while deep learning uses multiple hidden layers to extract representations of hierarchical features [76].

Figure 10

Explanation of the difference between a linear regression model and a simple learning model: (a) linear model regression model; (b) simplified deep learning model [77].

Figure 11

Process of segmenting medical images using deep learning [78].

Figure 12

The SegNet CNN architecture is depicted in this diagram. SegNet is made up of two networks: an encoder and a decoder. This network’s final output is high-resolution pixel-by-pixel tissue categorization [83].

Figure 13

Schematic representation of the workflow of [87] approach.

Figure 14

Demonstration of the number of papers reviewed for each method in KOA studies.