Two hundred 3D pre- and post-implant prostate brachytherapy MRIs were acquired with a T2w sequence, a T2/T1w sequence, or a T1w sequence. One hundred twenty DL models were trained to segment the prostate, seminal vesicles (SV), external urinary sphincter (EUS), rectum, and bladder using the MRIs acquired with T2w and T2/T1w image contrast. The DL models consisted of 18 fully convolutional networks (FCNs) with different convolutional encoders. Both 2D and 3D U-Net FCNs were constructed for comparison. Six objective functions were investigated: cross-entropy, Jaccard distance, focal loss, and 3 variations of Tversky distance. The performance of the models was compared using similarity metrics including pixel accuracy, Jaccard index, Dice similarity coefficient (DSC), 95% Hausdorff distance, relative volume difference, Matthews correlation coefficient, precision, recall, and average symmetric surface distance. We selected the highest-performing architecture and investigated how the amount of training data, use of skip connections, and data augmentation affected segmentation performance. Additionally, we investigated whether segmentation on the T1w MRIs was possible with FCNs trained on only T2w and T2/T1w image contrast.
Overall, an FCN with a DenseNet201 encoder trained via cross-entropy minimization yielded the highest combined segmentation performance. For the 53 3D test MRIs acquired with T2w or T2/T1w image contrast, the DSCs of the prostate, EUS, SV, rectum, and bladder were 0.90±0.04, 0.70±0.15, 0.80±0.12, 0.91±0.06, 0.96±0.04, respectively, after model fine-tuning. For the 5 T1w images, the DSCs of these organs were 0.82±0.07, 0.17±0.15, 0.46±0.21, 0.87±0.06, and 0.88±0.05, respectively.
Machine segmentation of the prostate and surrounding anatomy on 3D MRIs acquired with different pulse sequences for MARS LDR prostate brachytherapy is possible with a single FCN.
Copyright © 2020 Elsevier Inc. All rights reserved.