Room: AAPM ePoster Library
Purpose: This study aims to develop a deep learning-based approach for automatic intraprostatic lesion segmentation to enhance cancer control by dose escalation.
Methods: According to the Prostate Imaging Reporting and Data System (PI-RADS), the contours of prostates and suspicious lesions with PI-RADS score = 4 were manually delineated on axial T2-weighted (T2W)-, apparent diffusion coefficient (ADC)- and high b-value diffusion-weighted imaging (DWI) images collected from our institution, to provide the ground truth data. Then a Branch-UNet (B-UNet) was developed, in which multi-parametric magnetic resonance imaging (mpMRI) images were processed separately in different encoder paths and their high-level features were fused into the decoder path. The two adjacent image slices were also input into different sub-branches to take the contour consistency into account. The deep supervision strategy was integrated into the network to speed up the convergence. The probability maps of the background, normal prostate and lesion were output, and the performance of the network was evaluated using the Dice similarity coefficient (DSC) as the main metric.
Results: The current database consists of 2923 mpMRI image slices from 136 patient cases, in which 162 lesions were contoured on 652 slices and all prostates were contoured on 1264 slices. As for the segmentation of lesions in the testing set, B-UNet could achieve a per case DSC of 0.6247, specificity of 0.9994, sensitivity of 0.6852; and global DSC of 0.7189, specificity of 0.9994, sensitivity of 0.7324. The performance of the network is partly limited by some misleading benign prostatic hyperplasia, which is also a recognized limitation of mpMRI diagnosis.
Conclusion: The proposed deep learning-based approach segmented suspicious lesions automatically with the mpMRI images, and has been applied in our clinical procedures for intensifying dose to MRI lesions during brachytherapy procedures. The database will keep growing to improve the performance of the B-UNet.