Pelvic U-Net: multi-label semantic segmentation of pelvic organs at risk for radiation therapy anal cancer patients using a deeply supervised shuffle attention convolutional neural network.

BACKGROUND: Delineation of organs at risk (OAR) for anal cancer radiation therapy treatment planning is a manual and time-consuming process. Deep learning-based methods can accelerate and partially automate this task. The aim of this study was to develop and evaluate a deep learning model for automated and improved segmentations of OAR in the pelvic region. METHODS: A 3D, deeply supervised U-Net architecture with shuffle attention, referred to as Pelvic U-Net, was trained on 143 computed tomography (CT) volumes, to segment OAR in the pelvic region, such as total bone marrow, rectum, bladder, and bowel structures. Model predictions were evaluated on an independent test dataset (n = 15) using the Dice similarity coefficient (DSC), the 95th percentile of the Hausdorff distance (HD95), and the mean surface distance (MSD). In addition, three experienced radiation oncologists rated model predictions on a scale between 1-4 (excellent, good, acceptable, not acceptable). Model performance was also evaluated with respect to segmentation time, by comparing complete manual delineation time against model prediction time without and with manual correction of the predictions. Furthermore, dosimetric implications to treatment plans were evaluated using different dose-volume histogram (DVH) indices. RESULTS: Without any manual corrections, mean DSC values of 97%, 87% and 94% were found for total bone marrow, rectum, and bladder. Mean DSC values for bowel cavity, all bowel, small bowel, and large bowel were 95%, 91%, 87% and 81%, respectively. Total bone marrow, bladder, and bowel cavity segmentations derived from our model were rated excellent (89%, 93%, 42%), good (9%, 5%, 42%), or acceptable (2%, 2%, 16%) on average. For almost all the evaluated DVH indices, no significant difference between model predictions and manual delineations was found. Delineation time per patient could be reduced from 40 to 12 min, including manual corrections of model predictions, and to 4 min without corrections. CONCLUSIONS: Our Pelvic U-Net led to credible and clinically applicable OAR segmentations and showed improved performance compared to previous studies. Even though manual adjustments were needed for some predicted structures, segmentation time could be reduced by 70% on average. This allows for an accelerated radiation therapy treatment planning workflow for anal cancer patients.

Development and evaluation of a deep learning based artificial intelligence for automatic identification of gold fiducial markers in an MRI-only prostate radiotherapy workflow.

Identification of prostate gold fiducial markers in magnetic resonance imaging (MRI) images is challenging when CT images are not available, due to misclassifications from intra-prostatic calcifications. It is also a time consuming task and automated identification methods have been suggested as an improvement for both objectives. Multi-echo gradient echo (MEGRE) images have been utilized for manual fiducial identification with 100% detection accuracy. The aim is therefore to develop an automatic deep learning based method for fiducial identification in MRI images intended for MRI-only prostate radiotherapy. MEGRE images from 326 prostate cancer patients with fiducials were acquired on a 3T MRI, post-processed with N4 bias correction, and the fiducial center of mass (CoM) was identified. A 9 mm radius sphere was created around the CoM as ground truth. A deep learning HighRes3DNet model for semantic segmentation was trained using image augmentation. The model was applied to 39 MRI-only patients and 3D probability maps for fiducial location and segmentation were produced and spatially smoothed. In each of the three largest probability peaks, a 9 mm radius sphere was defined. Detection sensitivity and geometric accuracy was assessed. To raise awareness of potential false findings a 'BeAware' score was developed, calculated from the total number and quality of the probability peaks. All datasets, annotations and source code used were made publicly available. The detection sensitivity for all fiducials were 97.4%. Thirty-six out of thirty-nine patients had all fiducial markers correctly identified. All three failed patients generated a user notification using the BeAware score. The mean absolute difference between the detected fiducial and ground truth CoM was 0.7 ± 0.9 [0 3.1] mm. A deep learning method for automatic fiducial identification in MRI images was developed and evaluated with state-of-the-art results. The BeAware score has the potential to notify the user regarding patients where the proposed method is uncertain.

Dosimetric effects of adaptive prostate cancer radiotherapy in an MR-linac workflow.

BACKGROUND: The purpose was to evaluate the dosimetric effects in prostate cancer treatment caused by anatomical changes occurring during the time frame of adaptive replanning in a magnetic resonance linear accelerator (MR-linac) workflow. METHODS: Two MR images (MR1 and MR2) were acquired with 30 min apart for each of the 35 patients enrolled in this study. The clinical target volume (CTV) and organs at risk (OARs) were delineated based on MR1. Using a synthetic CT (sCT), ultra-hypofractionated VMAT treatment plans were created for MR1, with three different planning target volume (PTV) margins of 7 mm, 5 mm and 3 mm. The three treatment plans of MR1, were recalculated onto MR2 using its corresponding sCT. The dose distribution of MR2 represented delivered dose to the patient after 30 min of adaptive replanning, omitting motion correction before beam on. MR2 was registered to MR1, using deformable registration. Using the inverse deformation, the structures of MR1 was deformed to fit MR2 and anatomical changes were quantified. For dose distribution comparison the dose distribution of MR2 was warped to the geometry MR1. RESULTS: The mean center of mass vector offset for the CTV was 1.92 mm [0.13 - 9.79 mm]. Bladder volume increase ranged from 12.4 to 133.0% and rectum volume difference varied between -10.9 and 38.8%. Using the conventional 7 mm planning target volume (PTV) margin the dose reduction to the CTV was 1.1%. Corresponding values for 5 mm and 3 mm PTV margin were 2.0% and 4.2% respectively. The dose to the PTV and OARs also decreased from D1 to D2, for all PTV margins evaluated. Statistically significant difference was found for CTV Dmin between D1 and D2 for the 3 mm PTV margin (p < 0.01). CONCLUSIONS: A target underdosage caused by anatomical changes occurring during the reported time frame for adaptive replanning MR-linac workflows was found. Volume changes in both bladder and rectum caused large prostate displacements. This indicates the importance of thorough position verification before treatment delivery and that the workflow needs to speed up before introducing margin reduction.