A radiomic signature for detecting prostate cancer with multi-parametric MRI

Purpose: To detect prostate cancer (PCa) in the presence of clinical suspicion by exploiting a radiomic model on multi-parametric MRI sequences in order to support radiologists in daily clinical diagnostic pathway. Materials and Methods: 35 patients with clinical suspicion of PCa underwent 3T-mpMRI and TRUS-biopsy. 10 of them resulted healthy, 25 showed some evidences: 3 had lesions at mpMRI, with ASAP or PIN at biopsy, 22 had clear PCa (True Positives). Two experienced radiologists in consensus contoured all prostate slices and lesions of the 25 patients. Then, 84 radiomic features were extracted from b-2000 DWI sequences of the whole prostate. To increase the number of healthy samples, the features were also extracted outside the annotated regions, where no-clear evidence was detected, finally achieving 35 supposed-healthy samples (True Negatives). To derive the radiomic signature, LASSO with 3-fold Cross Validation and bootstrapping was performed on training and test sets. Results: The skewness of local entropy and the entropy of local coefficient of variation are differently representative of the heterogeneity of local tumour cellularity and define the radiomic signature. AUCmean of the trainings was 0.86(95%CI, 0.69-0.95) with specificity=93%, sensitivity=0.68%. On the test sets, the signature allows reaching AUCmean=0.85 (95%CI, 0.55-0.98), specificity=88%, sensitivity=0.70%. Conclusions: The radiomic features highlight properties not detectable by radiologists at visual inspection. In fact, the outcomes prove that effectiveness in PCa detection depends on the local variability of water restriction, rather than on the absolute DWI values. Limitations: Including not only clear healthy subjects in the True Negative class may lead to miss clear PCa patients. Nevertheless, the sensitivity might only increase, yet more confirming the potentiality of radiomic signature in PCa detection.