Utility of the K-Means Clustering Algorithm in Differentiating Apparent Diffusion Coefficient Values of Benign and Malignant Neck Pathologies

Does the K-means algorithm do a better job of differentiating benign and malignant neck pathologies compared to only mean ADC? The objective of our study was to analyze the differences between ADC partitions to evaluate whether the K-means technique can be of additional benefit to whole-lesion mean ADC alone in distinguishing benign and malignant neck pathologies. MR imaging studies of 10 benign and 10 malignant proved neck pathologies were postprocessed on a PC by using in-house software developed in Matlab. Two neuroradiologists manually contoured the lesions, with the ADC values within each lesion clustered into 2 (low, ADC-ADC(L); high, ADC-ADC(H)) and 3 partitions (ADC(L); intermediate, ADC-ADC(I); ADC(H)) by using the K-means clustering algorithm. An unpaired 2-tailed Student t test was performed for all metrics to determine statistical differences in the means of the benign and malignant pathologies. A statistically significant difference between the mean ADC(L) clusters in benign and malignant pathologies was seen in the 3-cluster models of both readers (P = .03 and .022, respectively) and the 2-cluster model of reader 2 (P = .04), with the other metrics (ADC(H), ADC(I); whole-lesion mean ADC) not revealing any significant differences. ROC curves demonstrated the quantitative differences in mean ADC(H) and ADC(L) in both the 2- and 3-cluster models to be predictive of malignancy (2 clusters: P = .008, area under curve = 0.850; 3 clusters: P = .01, area under curve = 0.825). The K-means clustering algorithm that generates partitions of large datasets may provide a better characterization of neck pathologies and may be of additional benefit in distinguishing benign and malignant neck pathologies compared with whole-lesion mean ADC alone.