Detection of dominant intraprostatic lesions in patients with prostate cancer using an artificial neural network and MR multimodal radiomics analysis

Document Type

Conference Proceeding

Publication Date


Publication Title

Int J Radiat Oncol Biol Phys


Purpose/Objective(s): Detection and classification of Dominant Intraprostatic Lesions (DILs) play an important role in diagnosis and radiation treatment response assessment of patients with prostate cancer (PCa). The aim of this pilot study was to construct an adaptive model for characterization and detection of DILs from normal tissue using radiomics features extracted from T2-weighted images and Apparent Diffusion Coefficient (ADC) maps. Purpose/Objective(s): Twenty-one patients with evidence of PCa with no prior treatment underwent MRI study. An ultrasound guided needle biopsy was performed to confirm the diagnosis. Two image modalities were acquired from the pelvis of all patients using a 3.0 T MR scanner: Axial T2 Weighted (T2W) fast spin echo (TE/TR = 4389/110 ms, FA = 90°, voxel size = 0.42 X 0.42 X 2.4 mm3) and axial diffusion weighted (DW) imaging (TE/TR = 4000/85 ms, FA = 90°, voxel size = 1.79 X 1.79 X 0.56 mm3, with b-values = 0 and 1000 [sec-mm-2] for constructing ADC maps). Using the diagnostic report of each PCa patient, a set of ROIs delineating DILs and normal tissues were drawn on each image modality. One hundred sixty-five radiomics features were extracted from tumor and normal volumes. A Partial Least Square Correlation (PLSC) along with one-way ANOVA were recruited to identify the most discriminant radiomics features (DILs versus normal) from multimodal information (T1WI and ADC). An artificial neural networks (ANN) was constructed based on the optimal feature set to classify the DILs and normal tissue. Using Leave-One-Out Cross Validation (LOOCV) techniques, the ANN was trained, optimized, and finally evaluated. Results: Among 165 radiomics features, 8 features (2 features from Two-Dimensional Wavelet Transform, 4 features from Two-Dimensional Gabor Filter, 2 features from Gray Level Co-occurrence Matrix) were found to be significantly different between the DILs and normal tissue in presence of multimodal information. Using the 8 radiomics features as input, after training and optimization (with 8, 5, and 1 neurons in its input, hidden and output layers, termination error = 0.022) of the ANN with LOOCV, it was able to differentiate the DILs and normal groups with predictive power (Az Test) of 84%. When the training vector was randomly permuted 1000 times, the permutation-invariant efficiency of the ANN was 4.5%. Conclusion: This pilot study demonstrates the feasibility of combining radiomics analysis using multimodal MR information and adaptive model to detect DILs in patients with prostate cancer. The study is limited by the number of patients, which can impact the optimal features selected, and also might render a predictive model susceptible to Type II errors. Additionally, the radiomics features selected from multimodal MRI might be impacted by the intensities and contrast of the T2WI and ADC maps. These factors, and the incorporation of additional MR modalities along with pathological-based information into the adaptive model, are being investigated.





First Page


Last Page


This document is currently not available here.