Uncertainty estimation using a 3D probabilistic U-Net for segmentation with small radiotherapy clinical trial datasets.

BACKGROUND AND OBJECTIVES: Bio-medical image segmentation models typically attempt to predict one segmentation that resembles a ground-truth structure as closely as possible. However, as medical images are not perfect representations of anatomy, obtaining this ground truth is not possible. A surrogate commonly used is to have multiple expert observers define the same structure for a dataset. When multiple observers define the same structure on the same image there can be significant differences depending on the structure, image quality/modality and the region being defined. It is often desirable to estimate this type of aleatoric uncertainty in a segmentation model to help understand the region in which the true structure is likely to be positioned. Furthermore, obtaining these datasets is resource intensive so training such models using limited data may be required. With a small dataset size, differing patient anatomy is likely not well represented causing epistemic uncertainty which should also be estimated so it can be determined for which cases the model is effective or not. METHODS: We use a 3D probabilistic U-Net to train a model from which several segmentations can be sampled to estimate the range of uncertainty seen between multiple observers. To ensure that regions where observers disagree most are emphasised in model training, we expand the Generalised Evidence Lower Bound (ELBO) with a Constrained Optimisation (GECO) loss function with an additional contour loss term to give attention to this region. Ensemble and Monte-Carlo dropout (MCDO) uncertainty quantification methods are used during inference to estimate model confidence on an unseen case. We apply our methodology to two radiotherapy clinical trial datasets, a gastric cancer trial (TOPGEAR, TROG 08.08) and a post-prostatectomy prostate cancer trial (RAVES, TROG 08.03). Each dataset contains only 10 cases each for model development to segment the clinical target volume (CTV) which was defined by multiple observers on each case. An additional 50 cases are available as a hold-out dataset for each trial which had only one observer define the CTV structure on each case. Up to 50 samples were generated using the probabilistic model for each case in the hold-out dataset. To assess performance, each manually defined structure was matched to the closest matching sampled segmentation based on commonly used metrics. RESULTS: The TOPGEAR CTV model achieved a Dice Similarity Coefficient (DSC) and Surface DSC (sDSC) of 0.7 and 0.43 respectively with the RAVES model achieving 0.75 and 0.71 respectively. Segmentation quality across cases in the hold-out datasets was variable however both the ensemble and MCDO uncertainty estimation approaches were able to accurately estimate model confidence with a p-value < 0.001 for both TOPGEAR and RAVES when comparing the DSC using the Pearson correlation coefficient. CONCLUSIONS: We demonstrated that training auto-segmentation models which can estimate aleatoric and epistemic uncertainty using limited datasets is possible. Having the model estimate prediction confidence is important to understand for which unseen cases a model is likely to be useful.

Evaluating the relationship between contouring variability and modelled treatment outcome for prostate bed radiotherapy.

Objectives.Contouring similarity metrics are often used in studies of inter-observer variation and automatic segmentation but do not provide an assessment of clinical impact. This study focused on post-prostatectomy radiotherapy and aimed to (1) identify if there is a relationship between variations in commonly used contouring similarity metrics and resulting dosimetry and (2) identify the variation in clinical target volume (CTV) contouring that significantly impacts dosimetry.Approach.The study retrospectively analysed CT scans of 10 patients from the TROG 08.03 RAVES trial. The CTV, rectum, and bladder were contoured independently by three experienced observers. Using these contours reference simultaneous truth and performance level estimation (STAPLE) volumes were established. Additional CTVs were generated using an atlas algorithm based on a single benchmark case with 42 manual contours. Volumetric-modulated arc therapy (VMAT) treatment plans were generated for the observer, atlas, and reference volumes. The dosimetry was evaluated using radiobiological metrics. Correlations between contouring similarity and dosimetry metrics were calculated using Spearman coefficient (Γ). To access impact of variations in planning target volume (PTV) margin, the STAPLE PTV was uniformly contracted and expanded, with plans created for each PTV volume. STAPLE dose-volume histograms (DVHs) were exported for plans generated based on the contracted/expanded volumes, and dose-volume metrics assessed.Mainresults. The study found no strong correlations between the considered similarity metrics and modelled outcomes. Moderate correlations (0.5 <Γ< 0.7) were observed for Dice similarity coefficient, Jaccard, and mean distance to agreement metrics and rectum toxicities. The observations of this study indicate a tendency for variations in CTV contraction/expansion below 5 mm to result in minor dosimetric impacts.Significance. Contouring similarity metrics must be used with caution when interpreting them as indicators of treatment plan variation. For post-prostatectomy VMAT patients, this work showed variations in contours with an expansion/contraction of less than 5 mm did not lead to notable dosimetric differences, this should be explored in a larger dataset to assess generalisability.

Clinical target volume delineation quality assurance for MRI-guided prostate radiotherapy using deep learning with uncertainty estimation.

BACKGROUND AND PURPOSE: Previous studies on automatic delineation quality assurance (QA) have mostly focused on CT-based planning. As MRI-guided radiotherapy is increasingly utilized in prostate cancer treatment, there is a need for more research on MRI-specific automatic QA. This work proposes a clinical target volume (CTV) delineation QA framework based on deep learning (DL) for MRI-guided prostate radiotherapy. MATERIALS AND METHODS: The proposed workflow utilized a 3D dropblock ResUnet++ (DB-ResUnet++) to generate multiple segmentation predictions via Monte Carlo dropout which were used to compute an average delineation and area of uncertainty. A logistic regression (LR) classifier was employed to classify the manual delineation as pass or discrepancy based on the spatial association between the manual delineation and the network's outputs. This approach was evaluated on a multicentre MRI-only prostate radiotherapy dataset and compared with our previously published QA framework based on AN-AG Unet. RESULTS: The proposed framework achieved an area under the receiver operating curve (AUROC) of 0.92, a true positive rate (TPR) of 0.92 and a false positive rate of 0.09 with an average processing time per delineation of 1.3 min. Compared with our previous work using AN-AG Unet, this method generated fewer false positive detections at the same TPR with a much faster processing speed. CONCLUSION: To the best of our knowledge, this is the first study to propose an automatic delineation QA tool using DL with uncertainty estimation for MRI-guided prostate radiotherapy, which can potentially be used for reviewing prostate CTV delineation in multicentre clinical trials.

Standardising Breast Radiotherapy Structure Naming Conventions: A Machine Learning Approach.

In progressing the use of big data in health systems, standardised nomenclature is required to enable data pooling and analyses. In many radiotherapy planning systems and their data archives, target volumes (TV) and organ-at-risk (OAR) structure nomenclature has not been standardised. Machine learning (ML) has been utilised to standardise volumes nomenclature in retrospective datasets. However, only subsets of the structures have been targeted. Within this paper, we proposed a new approach for standardising all the structures nomenclature by using multi-modal artificial neural networks. A cohort consisting of 1613 breast cancer patients treated with radiotherapy was identified from Liverpool & Macarthur Cancer Therapy Centres, NSW, Australia. Four types of volume characteristics were generated to represent each target and OAR volume: textual features, geometric features, dosimetry features, and imaging data. Five datasets were created from the original cohort, the first four represented different subsets of volumes and the last one represented the whole list of volumes. For each dataset, 15 sets of combinations of features were generated to investigate the effect of using different characteristics on the standardisation performance. The best model reported 99.416% classification accuracy over the hold-out sample when used to standardise all the nomenclatures in a breast cancer radiotherapy plan into 21 classes. Our results showed that ML based automation methods can be used for standardising naming conventions in a radiotherapy plan taking into consideration the inclusion of multiple modalities to better represent each volume.

Automatic radiotherapy delineation quality assurance on prostate MRI with deep learning in a multicentre clinical trial.

Volume delineation quality assurance (QA) is particularly important in clinical trial settings where consistent protocol implementation is required, as outcomes will affect future as well current patients. Currently, where feasible, this is conducted manually, which is time consuming and resource intensive. Although previous studies mostly focused on automating delineation QA on CT, magnetic resonance imaging (MRI) is being increasingly used in radiotherapy treatment. In this work, we propose to perform automatic delineation QA on prostate MRI for both the clinical target volume (CTV) and organs-at-risk (OARs) by using delineations generated by 3D Unet variants as benchmarks for QA. These networks were trained on a small gold standard atlas set and applied on a multicentre radiotherapy clinical trial dataset to generate benchmark delineations. Then, a QA stage was designed to recommend 'pass', 'minor correction' and 'major correction' for each manual delineation in the trial set by thresholding its Dice similarity coefficient to the network generated delineation. Among all 3D Unet variants explored, the Unet with anatomical gates in an AtlasNet architecture performed the best in delineation QA, achieving an area under the receiver operating characteristics curve of 0.97, 0.92, 0.89 and 0.97 for identifying unacceptable (major correction) delineations with a sensitivity of 0.93, 0.73, 0.74 and 0.90 at a specificity of 0.93, 0.86, 0.86 and 0.95 for bladder, prostate CTV, rectum and gel spacer respectively. To the best of our knowledge, this is the first study to propose automated delineation QA for a multicentre radiotherapy clinical trial with treatment planning MRI. The methods proposed in this work can potentially improve the accuracy and consistency of CTV and OAR delineation in radiotherapy treatment planning.

Contour variation is a primary source of error when delivering post prostatectomy radiotherapy: Results of the Trans-Tasman Radiation Oncology Group 08.03 Radiotherapy Adjuvant Versus Early Salvage (RAVES) benchmarking exercise.

INTRODUCTION: Variation in target volume delineation from clinical trial protocols has been shown to contribute to poorer patient outcomes. A clinical trial quality assurance framework can support compliance with trial protocol. Results of the TROG 08.03 RAVES benchmarking exercise considering variation from protocol, inter-observer variability and impact on dosimetry are reported in this paper. METHODS: Clinicians were required to contour and plan a benchmarking case according to trial protocol. Geometric pjmirometers including volume, Hausdorff Distance, Mean Distance to Agreement and DICE similarity coefficient were analysed for targets and organs at risk. Submitted volumes were compared to a STAPLE and consensus 'reference' volume for each structure. Dosimetric analysis was performed using dose volume histogram data. RESULTS: Benchmarking exercise submissions were received from 96 clinicians. In total 205 protocol variations were identified. The most common variation was inadequate contouring of the CTV in 84/205 (41%). The CTV volume ranged from 65.3 to 193.1 cm3 with a median of 113.2 cm3 . The most common dosimetric protocol variation related to rectal dosimetry. The mean submitted rectal volume receiving 40 Gy and 60 Gy, respectively, was 56.14% ± 5.55% and 30.25% ± 6.15%. When corrected to the protocol defined length the mean rectal volume receiving 40 Gy was 60.8% ± 7.92%, while the volume receiving 60 Gy was 33.86% ± 8.21%. CONCLUSION: Variations from protocol were found in the RAVES benchmarking exercise, most notably in CTV and rectum delineation. Inter-observer variability was evident. Incorrect delineation of the rectum impacted on dosimetric compliance with protocol.

Superior target volume and organ stability with the use of endorectal balloons in post-prostatectomy radiotherapy.

INTRODUCTION: We investigated the endorectal balloon (ERB) as a method to improve post-prostatectomy clinical target volume (CTV) stability. METHODS: Seventy cone-beam CT (CBCT) obtained during radiotherapy treatment from seven patients treated with an ERB and 68 CBCT from seven patients treated without an ERB were contoured according to published guidelines. CTV was subdivided into superior and inferior CTV; whole rectal volume was subdivided into superior and inferior rectum and anal volume. Concordance index (CI) of CBCT treatment volumes compared with planning volumes was calculated and displacements were measured. RESULTS: Whole rectal, superior and inferior rectum and anal CI were significantly improved with the ERB by 21%, 17%, 26% and 17% respectively (P < 0.0001). Overall CTV and inferior CTV CI was improved by 4% with the ERB (overall CTV P = 0.021; Inferior CTV P < 0.0001). In the ERB cohort, average displacement for superior CTV was 0.37 cm anterior-posterior (AP) and 0.10 cm left-right (LR). Average standard deviation was 0.27 cm AP and 0.11 cm LR. Inferior CTV average displacement was 0.11 cm AP and 0.02 cm LR. Average standard deviation was 0.11 cm AP and 0.02 cm LR. In the non-ERB cohort, average displacement for superior CTV was 0.43 cm AP and 0.10 mm left-right (LR). Average standard deviation was 0.45 cm AP and 0.13 cm LR. Inferior CTV average displacement was 0.16 cm AP and 0.01 cm LR. Average standard deviation was 0.17 cm AP and 0.03 cm LR. There was no statistically significant impact of bladder filling on CTV CI in ERB patients (P = 0.551) as opposed to non-ERB patients (P = 0.0421). CONCLUSION: ERBs in the post-prostatectomy setting resulted in increased rectal and CTV stability while negating the effects of bladder filling on CTV stability.

Endorectal balloons in the post prostatectomy setting: do gains in stability lead to more predictable dosimetry?

PURPOSE: To perform a comparative study assessing potential benefits of endorectal-balloons (ERB) in post-prostatectomy patients. METHOD AND MATERIALS: Ten retrospective post-prostatectomy patients treated without ERB and ten prospective patients treated with the ERB in situ were recruited. All patients received IMRT and IGRT using kilovoltage cone-beam computed tomography (kVCBCT). kVCBCT datasets were registered to the planning dataset, recontoured and the original plan recalculated on the kVCBCTs to recreate anatomical conditions during treatment. The imaging, structure and dose data were imported into in-house software for the assessment of geometric variation and cumulative equivalent uniform dose (EUD) in the two groups. RESULTS: The difference in location (ΔCOV) for the bladder between planning and each CBCT was similar for each group. The range of mean ΔCOV for the rectum was 0.15-0.58 cm and 0.15-0.59 cm for the non-ERB and ERB groups. For superior-CTV and inferior-CTV the difference between planned and delivered D95% (mean ± SD) for the non-ERB group was 2.1 ± 6.0 Gy and -0.04 ± 0.20 Gy. While for the ERB group the difference in D95% was 8.7 ± 12.6 Gy and 0.003 ± 0.104 Gy. CONCLUSIONS: The use of ERBs in the post-prostatectomy setting did improve geometric reproducibility of the target and surrounding normal tissues, however no improvement in dosimetric stability was observed for the margins employed.