A machine learning approach for distinguishing uterine sarcoma from leiomyomas based on perfusion weighted MRI parameters

•None of PWI parameters differed significantly between uterine sarcoma and leiomyomas.•Using a machine-learning method, PWI parameters were aggregated.•Computer was 92% accurate in uterine sarcoma/ leiomyomas classification. To propose a computer-assisted method for distinguishing uterine sarcoma from leiomyomas based on perfusion weighted magnetic resonance imaging (PWI). Forty-two women confirmed to have a total of 60 masses (10 uterine sarcomas and 50 benign leiomyomas) were included. The reference diagnosis was based on postoperative histopathological examination. All women underwent the standard MRI protocol with 3-Tesla MR imager (Magnetom Trio, Siemens, Erlangen, Germany) for assessment of myometrial masses, followed by PWI. For each mass, two regions of interest (ROI) were outlined manually by an experienced radiologist; one (ROIL) represented the entire tumor while the other (ROIs) was placed on the area of the lesion with the most marked contrast enhancement. Two additional ROIs with diameters similar to ROIs (3.0 to 3.1 mm) were placed on psoas muscle (ROIP) and myometrium (ROIM) in order to provide baselines for comparisons. The obtained ROIs of PWI images were then analyzed using the DCE Tool plug-in (version 2.0SP1) within ClearCanvas (Toronto, Ontario, Canada) framework. The DCE Tool provides seven parameters (Ktrans, kep, Vb, IAUC, initial slope, peak, the mean squared error) for modelling contrast uptake within an ROI using the modified Tofts model. Parameters extracted from the ROIs were fed into a decision tree ensemble, which classified the corresponding lesions either as malignant or benign. The leave-one-out cross validation (LOOCV) was utilized to evaluate the performance of the classifier. None of the parameters extracted from ROIL or ROIs differed significantly between uterine sarcoma and benign leiomyomas (all p > 0.05). The overall accuracy of 66.7% was obtained by feeding seven parameters extracted from ROIL to the classifier. When 21 features extracted from ROIL, ROIM, and ROIP were fed into the classifier an accuracy of 91.7%, sensitivity of 100%, and specificity of 90% were achieved in the optimal operating point of classifier. Although none of the PWI parameters differed significantly between benign and malignant lesions, when the information provided by the extracted features was aggregated using a machine learning method, a promising discriminative power was obtained. This suggests that the proposed model for combining the