Objectives

Accurate differentiation of temporary vs. permanent changes occurring following irreversible electroporation (IRE) holds immense importance for the early assessment of ablative treatment outcomes. Here, we investigated the benefits of advanced statistical learning models for an immediate evaluation of therapeutic outcomes by interpreting quantitative characteristics captured with conventional MRI.

Methods

The preclinical study integrated twenty-six rabbits with anatomical and perfusion MRI data acquired with a 3T clinical MRI scanner. T1w and T2w MRI data were quantitatively analyzed, and forty-six quantitative features were computed with four feature extraction methods. The candidate key features were determined by graph clustering following the filtering-based feature selection technique, RELIEFF algorithm. Kernel-based support vector machines (SVM) and random forest (RF) classifiers interpreting quantitative features of T1w, T2w, and combination (T1w+T2w) MRI were developed for replicating the underlying characteristics of the tissues to distinguish IRE ablation regions for immediate assessment of treatment response. Accuracy, sensitivity, specificity, and area under the receiver operating characteristics curve were used to evaluate classification performance.

Results

Following the analysis of quantitative variables, three features were integrated to develop a SVM classification model, while five features were utilized for generating RF classifiers. SVM classifiers demonstrated detection accuracy of 91.06%, 96.15%, and 98.04% for individual and combination MRI data, respectively. Besides, RF classifiers obtained slightly lower accuracy compared to SVM which were 95.06%, 89.40%, and 94.38% respectively.

Conclusions

Quantitative models integrating structural characteristics of conventional T1w and T2w MRI data with statistical learning techniques identified IRE ablation regions allowing early assessment of treatment status.