Utility of the k-means clustering algorithm in differentiating apparent diffusion coefficient values of benign and...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning in Medicine

Background:

• Distinguishing benign from malignant neck pathologies is crucial for appropriate patient management.
• Apparent Diffusion Coefficient (ADC) measurements are used in MRI to assess tissue characteristics.
• Current methods using whole-lesion mean ADC may have limitations in differentiating certain pathologies.

Purpose of the Study:

• To evaluate if the K-means clustering algorithm improves the differentiation of benign and malignant neck pathologies compared to mean ADC alone.
• To analyze the differences in ADC partitions generated by K-means clustering.
• To assess the added benefit of K-means technique in distinguishing neck pathologies.

Main Methods:

• MR imaging of 20 neck pathologies (10 benign, 10 malignant) were analyzed.
• ADC values were clustered into 2 or 3 partitions using the K-means algorithm.
• Statistical analysis (Student t test, ROC curves) compared K-means partitions and mean ADC values.

Main Results:

• The mean ADC(L) clusters showed statistically significant differences between benign and malignant pathologies in 3-cluster models (P = .03, .022) and a 2-cluster model (P = .04).
• ROC curves indicated that quantitative differences in mean ADC(H) and ADC(L) were predictive of malignancy (2 clusters: AUC = 0.850; 3 clusters: AUC = 0.825).
• Whole-lesion mean ADC and other ADC partitions did not reveal significant differences.

Conclusions:

• K-means clustering algorithm can provide better characterization of neck pathologies by partitioning large datasets.
• The K-means technique offers additional benefit in distinguishing benign and malignant neck pathologies compared to whole-lesion mean ADC alone.
• This approach may enhance diagnostic accuracy in neck lesion assessment.