Enhanced human knee cartilage evaluation using reduced interactive segmentation model

Abstract

The purpose of this research is to design an enhanced human knee cartilage evaluation framework to detect cartilage thinning in the early Osteoarthritis (OA) disease. The existing research drawbacks include the absence of contrast enhancement model merely on region of interest, the low efficiency and tedious labelling processes in interactive segmentation model, and the lacking of a quantitative assessment in the segmentation model. In this research, we propose a quantitative assessment framework which consists of three phases: Phase 1 focuses on developing an explicit contrast enhancement model for knee images; Phase 2 focuses on developing a reduced interactive cartilage segmentation tool; Phase 3 focuses on formulating a cartilage quantitative measurement. The knee images tested in this research are provided by Osteoarthritis Initiative, given that the sample sizes used were 120, 30 and 20 slices in Phase 1, Phase 2 and Phase 3, respectively. The proposed Prominent Region of Interest Contrast Enhancement (PROICE) method outperformed in diverging the dynamic range of intensity distributed by the region of interest, resulting in noticeable distinctiveness between cartilages and unwanted background tissues. Compared with other existing enhancement methods, PROICE achieved the highest peak signal-to-noise ratio score of 23.80±1.16dB, structural similarity index of 0.86±0.02, low absolute mean error score of 3.88±2.92, and adequate enhancement measure of 17.47±0.74. It was then extended to Enhanced Approximate Non-Cartilage Labels (EANCAL) for the extraction of portions that contained critical information through an entropy filter. This research contributed to reduce human attention level in manual annotations, eventually increased the segmentation efficiency. The modified segmentation framework showed a significant reduction in the mean processing time to 45±4s, which was averaged of 80.25% and 82.25% shorter than manual segmentation for healthy knee cartilage segmentation and diseased knee cartilage segmentation respectively, that performed by two trained operators. In addition, EANCAL obtained an adequate inter-operator reliability score in healthy femoral cartilage (FC) and tibial cartilage (TC) (FC: 0.920±0.046;TC:0.912±0.044). Meanwhile, EANCAL remained competitive compared to the ANCAL method yet with fewer human attention level required, recorded with the highest intra-operator reproducibility score of 0.820±0.074 for operator 1; and 0.833±0.056 for operator 2. The cartilage segmentations were then evaluated with Regional Cartilage Normal thickness approximation (RCN-ta). The quantitative assessment model was validated with FDA-cleared DICOM software, revealed an acceptable error range of 0.135-0.214 mm. The inter-class correlation score and Pearson correlation obtained were ICC>0.94 and r>0.90, respectively. In a nutshell, the PROICE-enhanced images successfully overcome the background seed allocation issue and improved the segmentation model efficiency and segmentation reproducibility, thus yielding a promising cartilage quantitative assessment framework, which potentially assist the clinicians in diagnosis and treatment decision-making process.