Whole-tumor texture model based on diffusion kurtosis imaging for assessing cervical cancer: a preliminary study.

ABSTRACT

OBJECTIVES: To evaluate the diagnostic potential of diffusion kurtosis imaging (DKI) functional maps with whole-tumor texture analysis in differentiating cervical cancer (CC) subtype and grade. METHODS: Seventy-six patients with CC were enrolled. First-order texture features of the whole tumor were extracted from DKI and DWI functional maps, including apparent kurtosis coefficient averaged over all directions (MK), kurtosis along the axial direction (Ka), kurtosis along the radial direction (Kr), mean diffusivity (MD), fractional anisotropy (FA), and ADC maps, respectively. The Mann-Whitney U test and ROC curve were used to select the most representative texture features. Models based on each individual and combined functional maps were established using multivariate logistic regression analysis. Conventional parameters-the average values of ADC and DKI parameters derived from the conventional ROI method-were also evaluated. RESULTS: The combined model based on Ka, Kr, MD, and FA maps yielded the best diagnostic performance in discrimination of cervical squamous cell cancer (SCC) and cervical adenocarcinoma (CAC) with the highest AUC (0.932). Among individual functional map derived models, Kr map-derived model showed the best performance when differentiating tumor subtypes (AUC = 0.828). MK_90th percentile was useful for distinguishing high-grade and low-grade in SCC tumors with an AUC of 0.701. The average values of MD, FA, and ADC were significantly different between SCC and CAC, but no conventional parameters were useful for tumor grading. CONCLUSIONS: The whole-tumor texture analysis applied to DKI functional maps can be used for differential diagnosis of cervical cancer subtypes and grading SCC. KEY POINTS • The whole-tumor texture analysis applied to DKI functional maps allows accurate differential diagnosis of CC subtype and grade. • The combined model derived from multiple functional maps performs significantly better than the single models when differentiating tumor subtypes. • MK_90th percentile was useful for distinguishing poorly and well-/moderately differentiated SCC tumors with an AUC of 0.701.