Gaussian mixture model-based cluster analysis of apparent diffusion coefficient values: a novel approach to evaluate...

The purpose of our study was to perform Gaussian mixture model (GMM)-based cluster analysis of the apparent diffusion coefficient (ADC) data of patients with endometrioid carcinoma, and to evaluate the relationship between histological grade and the ratios of the different clusters in each patient.

This retrospective study enrolled 122 patients (training: n = 63; and validation: n = 59) imaged between May 2015 and February 2020. In the training cohort, manual segmentation was performed on the ADC maps to obtain the ADC data of each patient, and these ADC data were summated to obtain the "All-patient" ADC data. Cluster analysis (three clusters) was performed on this All-patient ADC data, and the ADC ranges of each cluster were defined as follows: cluster 1, 490-699 × 10^-6 mm²/s; cluster 2, 700-932 × 10^-6 mm²/s; and cluster 3, over 933 × 10^-6 mm²/s. In the training and validation cohorts, the ADC data of each patient was classified into three clusters according to these ADC ranges. The cluster ratios of each patient were calculated and compared with histological grade.

In the training cohort, a significant positive correlation was found between the cluster 1 ratio and histological grade (ρ = 0.34, p = 0.0059). The cluster 1 ratios of high-grade lesions (grade 3) were significantly higher than those of low-grade lesions (grades 1 and 2) (p = 0.0084). A similar significant positive correlation was found between the cluster 1 ratio and histological grade in the validation cohort (ρ = 0.44, p = 0.0006).

The cluster 1 ratio containing voxels with low ADC was significantly correlated with the histological grade of endometrioid carcinoma.

• We performed Gaussian mixture model (GMM)-based cluster analysis of the apparent diffusion coefficient (ADC) data of patients with endometrioid carcinoma. • The cluster 1 ratio, which included low ADC values, was significantly positive correlated with histological grade in the training and validation cohorts. • The GMM-based cluster analysis of voxel-based ADC data was effective for grading endometrioid carcinoma.