Usefulness of texture features of apparent diffusion coefficient maps in predicting chemoradiotherapy response in muscle-invasive bladder cancer

Objectives To examine the usefulness of the texture analysis (TA) of apparent diffusion coefficient (ADC) maps in predicting the chemoradiotherapy (CRT) response of muscle-invasive bladder cancer (MIBC). Methods We reviewed 45 MIBC patients who underwent cystectomy after CRT. CRT response was assessed through histologic evaluation of cystectomy specimens. Two radiologists determined the volume of interest for the index lesions on ADC maps of pretherapeutic 1.5-T MRI and performed TA using the LIFEx software. Forty-six texture features (TFs) were selected based on their contribution to the prediction of CRT sensitivity. To evaluate diagnostic performance, diagnostic models from the selected TFs were created using random forest (RF) and support vector machine (SVM), respectively. Results Twenty-three patients achieved pathologic complete response (pCR) to CRT. The feature selection identified first quartile ADC (Q1 ADC), gray-level co-occurrence matrix (GLCM) correlation, and GLCM homogeneity as important in predicting CRT response. Patients who achieved pCR showed significantly lower Q1 ADC and GLCM correlation values (0.66 x 10(-3) mm(2)/s and 0.53, respectively) than those who did not (0.81 x 10(-3) mm(2)/s and 0.70, respectively; p < 0.05 for both). The AUCs of the RF and SVM models incorporating the selected TFs were 0.82 (95% confidence interval [CI]: 0.67-0.97) and 0.96 (95% CI: 0.91-1.00), respectively, and the AUC of the SVM model was better than that of the mean ADC value (0.76, 95% CI: 0.61-0.90; p = 0.0037). Conclusion TFs can serve as imaging biomarkers in MIBC patients for predicting CRT sensitivity. TAs of ADC maps can potentially optimize patient selection for CRT.

Automated summary

Texture analysis of ADC maps can predict CRT sensitivity in MIBC patients, with Q1 ADC and GLCM identified as important texture features for predicting CRT response. The RF and SVM models incorporating the selected TFs showed promising diagnostic performance, suggesting potential for optimizing patient selection for CRT.