Pith. sign in

REVIEW 3 major objections 4 minor 42 references

AI-Assisted Decision-Making for Clinical Assessment of Auto-Segmented Contour Quality

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read A calibrated Bayesian ordinal classifier screens radiotherapy auto-contours with over 90% accuracy using only 30 manual labels, needing no ground-truth contours at run time.

desk verdict Clever integration of ordinal classification and MC-dropout uncertainty for contour QA, but the headline accuracy claims are not distinguished from the trivial majority-class rule on a badly imbalanced test set. read the letter →

arxiv 2505.00308 v2 pith:YBMQFGSR submitted 2025-05-01 cs.CV cs.AIstat.AP

classification cs.CVcs.AIstat.AP
keywords auto-contourqualityassessmentBayesianordinalclassificationuncertaintyquantificationonlineadaptiveradiotherapydeeplearningMonteCarlodropoutsurrogatelabelsQA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that auto-contour quality in radiotherapy can be assessed without ground-truth contours at run time, using a Bayesian ordinal classifier that outputs both a quality grade and an uncertainty for each 2D slice. The method is built to work when manual labels are absent, scarce, or plentiful: geometric surrogate labels cover the first case, transfer learning with as few as 30 labeled subjects covers the second, and direct supervision covers the third. A calibration step converts the uncertainty estimate into a threshold that meets a prespecified clinical accuracy target, and only predictions below that threshold are presented as confident. If the claim holds, clinicians in online adaptive radiotherapy could trust an automated screen for most auto-contours and spend manual review only on uncertain, rejected, or warning-triggering slices. The reported result is over 90% accuracy against manual labels after fine-tuning with 30 labeled subjects, with more than 98% of test slices accepted as confident and 93.6% of those confident calls correct.

What carries the argument

The load-bearing object is the Bayesian ordinal classifier, a ResNet-34 with dropout after each activation layer trained with the CORN conditional-probability loss; dropout turns it into a variational Bayesian network whose predictive distribution is approximated by 20 Monte Carlo forward passes. Ordinal coding is $y^*=y_1^*+y_2^*$, where the two binary decisions are 'at least Class 1' and 'at least Class 2', so the model respects that a rejected contour is worse than one needing major revision. Uncertainty is not the softmax entropy of a single pass but the total variance $\mathrm{Var}(y^*)$ decomposed through the law of total variance into within-weight and across-weight components, estimated from the Monte Carlo draws. The calibration step converts that variance into a threshold $\tau$ by sorting calibration slices by uncertainty and reading off the uncertainty level at which cumulative accuracy crosses the target; only slices with uncertainty below $\tau$ are reported as confident. Geometric surrogate labels, set by DSC, SDSC, and HD95 with the final class taken as the worst of the three metric ratings, provide the training signal when manual labels do not exist.

What would settle it

Apply the exact calibrated threshold (e.g., $\tau=0.33$ for the transfer-learning model) to an independent cohort in which acceptable contours are the minority rather than 93.1% of slices; if accuracy on accepted low-uncertainty cases drops below 90%, the distribution-transfer assumption is disproved. A reader can also compute the always-Class-2 accuracy on the current test set to see how much of the reported accuracy is explained by class imbalance alone.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an ordinal three-class Bayesian classifier, trained on CT slices overlaid with auto-contours, can be calibrated so that its confident predictions meet a prespecified clinical accuracy. Quality is coded as $y^*=y_1^*+y_2^*$, with $y_1^*=1$ meaning at least 'major revision' and $y_2^*=1$ meaning at least 'no/minor revision'; each binary event is modeled conditionally, and Monte Carlo dropout approximates the posterior over network weights. The variance of $y^*$ is used as predictive uncertainty, and a calibration curve of accuracy versus uncertainty yields a threshold for accepting predictions. In the limited-label scenario, pretraining on geometric surrogate labels (DSC, SDSC, HD95) and fine-tuning with 30 manually labeled subjects gives 92.9% overall accuracy, and applying the threshold to meet the 90% target gives 93.6% accuracy over 98.2% of the test slices. The same uncertainty measure tracks clinician disagreement, so low-uncertainty predictions are exactly the slices where the automated assessment can replace manual review.

Load-bearing premise

The calibration cohort and the future test cohort are assumed to produce the same distribution of contour-quality uncertainties, so a threshold tuned on 34 subjects keeps its 90% accuracy promise when applied to new patients.

Editorial extensions

If this is right

  • With zero manual labels, geometric surrogate labels plus data augmentation give a usable QA screen: 92.8% manual-label accuracy on 60.5% of test slices, enough to pre-screen contours where annotation is impossible.
  • With 20 to 30 manually labeled subjects used as fine-tuning, the model becomes confident on 96.5% to 98.7% of test slices and holds above 90% accuracy on those accepted predictions.
  • The same uncertainty threshold can be set to meet 80%, 85%, or 90% clinical accuracy targets, trading how many cases are auto-screened against how reliable the accepted calls are.
  • Model uncertainty tracks clinician disagreement: slices on which the three clinicians disagreed receive higher model uncertainty, so low-uncertainty predictions are the cases where a clinician can trust the assessment most.
  • Triggering a warning only when the model confidently disagrees with the clinician's 'no revision' judgment gives a decision-support loop that does not require ground-truth contours in deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Not developed in the paper: the same surrogate-label thresholds would have to be re-derived per organ, so the portability claim is about the BOC-plus-calibration machinery, not about the rectum-specific numbers.
  • An untested consequence: on a test set that is 93.1% Class 2, a trivial always-'acceptable' predictor already achieves 93.1% overall accuracy; the clinically decisive quantities are the recall and precision on Class 0 and Class 1 slices among confident predictions, which the paper does not isolate.
  • A likely operational extension: because uncertainty comes from 20 forward passes on 2D slices, the QA screen can run on the same machine as the auto-segmentation model and add only seconds per patient, making it usable inside an online adaptive session.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a Bayesian Ordinal Classification (BOC) model with Monte Carlo dropout uncertainty quantification to assess auto-segmented rectum contour quality in radiotherapy. The method is trained under three label-availability scenarios: geometric surrogate labels, transfer learning with limited manual labels, and fully supervised training. A calibration step selects an uncertainty threshold to meet a target accuracy (e.g., 90%), and only low-uncertainty predictions are accepted. The central reported result is that fine-tuning with 30 manually labeled subjects and calibrating on 34 subjects yields 92.9% overall accuracy and 93.6% accuracy on 98.2% of test cases with uncertainty below the calibrated threshold.

Significance. If the central claim were established, the paper would make a useful practical contribution: an uncertainty-gated QA model that avoids ground-truth contours and reduces manual review burden, with a formal variance decomposition for ordinal classification. The mathematical derivation in Section II.A.2 is sound, and the use of a separate calibration set and a held-out test set is methodologically appropriate. However, the reported test accuracies are not compared with the trivial majority-class baseline on a heavily imbalanced test set, and the central accuracy claim is therefore not supported by the evidence as presented.

major comments (3)
  1. [Section III.B, Table 6 and Appendix Table A1] The test set contains 93.1% Class 2 slices by clinician majority vote, so an always-Class-2 classifier achieves 93.1% accuracy. The transfer-learning model with 30 subjects achieves 92.9% overall accuracy, below that baseline, and the calibrated confident-subset accuracy of 93.6% on 98.2% of cases is only 0.5 percentage points above it. The manuscript reports neither a majority-class baseline nor a confusion matrix or per-class recall for the transfer-learning model, so the claimed >90% accuracy and the workload-reduction claim are not established. The reader cannot tell whether the uncertainty threshold selects genuinely reliable predictions or merely exploits the class prior.
  2. [Section II.A.3a and Table 6] The calibration procedure selects a threshold to achieve a target accuracy on calibration data and then reports accuracy on the test subset below that threshold. With 98.2% of test cases below the threshold, the confident subset is nearly the full test set, and its 93.6% accuracy is again indistinguishable from the 93.1% majority baseline. A valid demonstration requires reporting the confusion matrix on the confident subset, per-class recall and precision, and the accuracy of an always-Class-2 predictor restricted to the same subset. The stated assumption that the calibration and test distributions match is not verified, and the calibration set's class distribution is not reported.
  3. [Section II.A.2b and Fig. 4] The claim of a strong correlation between model uncertainty and manual-label uncertainty rests on a Spearman correlation of 1 computed from only three grouped points (full agreement, partial agreement, complete disagreement). With three groups, this only shows that the group means are monotonically ordered; it does not demonstrate a statistically meaningful correlation across individual cases. The authors should report the per-group distributions or a case-level correlation, and state the sample size in each agreement group.
minor comments (4)
  1. [Abstract] The phrase 'achieved over an accuracy of over 90%' is redundant and grammatically awkward; please revise.
  2. [Table 5] Table 5 is difficult to parse: the 'Uncertainty Level' column entries are not self-explanatory, and the rows for the 90% threshold do not clearly indicate which threshold applies. Add explicit column headers and complete cell entries.
  3. [Table 6] The table defines overall accuracy and confident-subset accuracy, but it would be much more informative to include per-class recall or a confusion matrix for the transfer-learning and from-scratch models, along with the calibration set's class distribution.
  4. [References and Fig. A1] Reference [13] is incomplete, as it lacks venue and year information, and the caption of Fig. A1 panel e says 'Class 0' while the text says clinicians assessed that case as 'Class 1'; please reconcile these details.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the calibration threshold is fit on a separate calibration cohort and the reported test accuracies are measured on a held-out test set; the majority-class baseline concern is a statistical weakness, not a circularity.

full rationale

The derivation chain is self-contained. The BOC posterior, predictive distribution, and uncertainty formulas in Sections II.A.1 and II.A.2 are derived from probability calculus rather than from the target result. MC dropout is justified by external references [22] and [31], and the CORN training procedure is adopted from external reference [23]. Surrogate labels are constructed from geometric thresholds in Table 1 and used for training; the model is then separately evaluated against clinician manual labels in Section III.A.2 and Table 4, so surrogate-label agreement is not presented as manual-label agreement. The uncertainty threshold is calibrated on a distinct calibration set (e.g., 34 subjects for transfer learning, 297 subjects for surrogate-label models) by reading the accuracy-versus-uncertainty curve, and the claimed 90%-plus accuracy figures, including 93.6% on 98.2% of test cases, are measured on the held-out 47-subject test set (Table 6, Fig. 5). Thus the prediction does not reduce by construction to the calibration fit. The strongest weakness in the paper is statistical rather than circular: Appendix Table A1 shows 93.1% of the 1,922 test slices are Class 2 by majority vote, and no majority-class baseline is reported, so the absolute accuracy numbers may not demonstrate skill beyond the class prior. That is a real evaluation-design concern, but it does not make the derivation equivalent to its inputs. The self-citations present ([11]-[14], [18]) are contextual uncertainty-quantification citations and are not load-bearing premises of the central claim.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims depend on hand-picked geometric thresholds for surrogate labels, calibrated uncertainty thresholds, and a distribution-match assumption between calibration and test sets. No new physical or conceptual entities are introduced.

free parameters (3)
  • Surrogate label thresholds (DSC, SDSC, HD95) = DSC 0.9/0.7; SDSC 0.9/0.7; HD95 2.5/6.0 mm; final class = max
    Table 1 defines these thresholds by hand without clinical validation. They determine the surrogate training labels, so the model's learned mapping depends on them.
  • Uncertainty thresholds (tau) = e.g., 0.24 (surrogate with augmentation, 90% requirement), 0.33 (transfer learning, 90% requirement)
    Calibrated on calibration data to meet accuracy requirements; then applied to test data. These are fitted values that vary by scenario and accuracy target.
  • MC dropout sample count T = 20
    Chosen based on Kwon et al. rather than tuned per problem; affects the stability of uncertainty estimates.
assumptions (4)
  • domain assumption Test data distribution matches calibration data distribution
    Section II.A.3a assumes that predictions with uncertainty below a threshold calibrated on calibration data will achieve the required accuracy on test data. If this fails, the confident-prediction accuracy is not guaranteed.
  • domain assumption Geometric surrogate labels correlate with clinical quality
    Section II.B.1b derives surrogate labels from DSC/SDSC/HD95 thresholds. The paper cites correlations of 0.25 to 0.59 between metrics and manual rankings, but the specific thresholds in Table 1 are not clinically validated.
  • standard math MC dropout provides a valid variational approximation of the Bayesian posterior
    Relies on Gal et al. and Gal and Ghahramani, a standard approximation with known limitations; the paper does not quantify the approximation error.
  • domain assumption Three-point ordinal scale captures clinically meaningful quality differences
    Used throughout; assumes the ordinal classes are distinct and that the asymmetry in misclassification risk is as described in the introduction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-Assisted Decision-Making for Clinical Assessment of Auto-Segmented Contour Quality." pith.science (2026). https://pith.science/paper/YBMQFGSR

@misc{pith2026250500308,
  author       = {Pith},
  title        = {Pith review of: AI-Assisted Decision-Making for Clinical Assessment of Auto-Segmented Contour Quality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YBMQFGSR}},
  note         = {Machine review of arXiv:2505.00308}
}
read the original abstract

Purpose: This study presents a Deep Learning (DL)-based quality assessment (QA) approach for evaluating auto-generated contours (auto-contours) in radiotherapy, with emphasis on Online Adaptive Radiotherapy (OART). Leveraging Bayesian Ordinal Classification (BOC) and calibrated uncertainty thresholds, the method enables confident QA predictions without relying on ground truth contours or extensive manual labeling. Methods: We developed a BOC model to classify auto-contour quality and quantify prediction uncertainty. A calibration step was used to optimize uncertainty thresholds that meet clinical accuracy needs. The method was validated under three data scenarios: no manual labels, limited labels, and extensive labels. For rectum contours in prostate cancer, we applied geometric surrogate labels when manual labels were absent, transfer learning when limited, and direct supervision when ample labels were available. Results: The BOC model delivered robust performance across all scenarios. Fine-tuning with just 30 manual labels and calibrating with 34 subjects yielded over 90% accuracy on test data. Using the calibrated threshold, over 93% of the auto-contours' qualities were accurately predicted in over 98% of cases, reducing unnecessary manual reviews and highlighting cases needing correction. Conclusion: The proposed QA model enhances contouring efficiency in OART by reducing manual workload and enabling fast, informed clinical decisions. Through uncertainty quantification, it ensures safer, more reliable radiotherapy workflows.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 31 canonical work pages

  1. [1]

    Geometric-based methods[3,4]: Most geometric metrics require two contour sets (a test contour and a reference contour) to benchmark and quantify similarity. For example, the Dice Similarity Coefficient (DSC) measures the overlap between entire areas, whereas the Surface Dice Similarity Coefficient (SDSC), Hausdorff Distance (HD), Mean Surface Distance (MS...

  2. [2]

    Database-driven methods: Altman et al. [5] proposed a framework to evaluate contour quality by analyzing historical data on size, shape, position, and other clinically relevant metrics, and then setting thresholds based on their distributions. However, constructing a sufficiently robust database for each organ and metric is challenging. Furthermore, the a...

  3. [3]

    acceptable

    Prediction-based methods: These leverage machine learning (ML) and deep learning (DL) models to predict contour quality. Traditional ML approaches typically employ geometric metrics (e.g., DSC, SDSC, HD and MSD) as input features to predict contour quality[6, 7]. Men et al.[8] employed a deep active learning framework in a lung cancer clinical trial, gene...

  4. [4]

    Let 𝒴 denote the label space of ordinal quality levels, 𝒴 = {𝑟0, 𝑟1, … , 𝑟𝑘, … 𝑟𝐾−1} with 𝑟0 = 0, 𝑟𝑘 = 𝑘, 𝑘 = 1,2, … , 𝐾 − 1, and 𝑟𝐾−1 > ⋯ > 𝑟0

    Deep Learning-based Bayesian Ordinal Classification (BOC) Model Let 𝒳 denote the input space consisting of all 2D CT slices together with their associated auto -contours. Let 𝒴 denote the label space of ordinal quality levels, 𝒴 = {𝑟0, 𝑟1, … , 𝑟𝑘, … 𝑟𝐾−1} with 𝑟0 = 0, 𝑟𝑘 = 𝑘, 𝑘 = 1,2, … , 𝐾 − 1, and 𝑟𝐾−1 > ⋯ > 𝑟0 . Given a training dataset 𝐷 = {𝑋(𝑖), 𝑦(𝑖)...

  5. [5]

    Uncertainty Quantification Method a) Uncertainty Estimated from the BOC Model To quantify prediction uncertainty, we consider the variance of 𝑦∗, 𝑉𝑎𝑟(𝑦∗) = 𝑉𝑎𝑟(∑𝑘=1 𝐾−1 𝑦𝑘 ∗) = 𝑉𝑎𝑟(𝑦1 ∗ + 𝑦2 ∗). The first two central moments of 𝑦∗ with respect to the distribution 𝑝(𝑦∗|𝑋∗, 𝑊) can be given by: 𝔼𝑝(𝑦∗|𝑋∗,𝑊)(𝑦∗) = 𝔼𝑝(𝑦∗|𝑋∗,𝑊)(𝑦1 ∗) + 𝔼𝑝(𝑦∗|𝑋∗,𝑊)(𝑦2 ∗) = ∑{𝑦1 ∗...

  6. [6]

    accuracy vs. uncertainty

    AI-Assisted Decision-making Strategies for Application a) Calibration Our BOC model integrated both estimated uncertainty and predicted quality in decision-making processes for auto-contour QA. We hypothesized that effective uncertainty quantification would demonstrate a direct correla- tion between decreased uncertainty and improved model performance. To...

  7. [7]

    The CT scans from these patients had slice thickness ranging from 1.5 to 3.0 mm, resolution of 0.9 to 1.5 mm, image size of 512x512 pixels, and 90 to 290 slices per scan

    Data a) Image Data In this single -institution study approved by our institutional review board (IRB), we curated a cohort of 1,328 prostate cancer patients from U niversity of Texas Southwestern Medical Center , each with at least one cleaned GT contour for the rectum, bladder or prostate. The CT scans from these patients had slice thickness ranging from...

  8. [8]

    The model was trained using the Adam optimizer with default hyperparameters ( β1 = 0.9, and β2 = 0.999) over 1 × 105 iterations, leveraging the dice loss function

    Experimental Design a) DL- based auto-segmentation model We trained an in-house 3D auto-segmentation model based on the open -source MONAI U-Net to generate auto- contours of the prostate, rectum and bladder for the downstream QA task. The model was trained using the Adam optimizer with default hyperparameters ( β1 = 0.9, and β2 = 0.999) over 1 × 105 iter...

Show all 42 references
  1. [9]

    As shown in Table 2, the ROC AUC values increased from 0.821 to 0.932 and from 0.778 to 0.808 for predicting Class 2 and Class 1 , respectively

    Uncertainty Guided Decision Making Overall Performance In testing samples, the model trained with surrogate labels achieved an overall accuracy of 70.8%, which increased to 79.2% with data augmentation. As shown in Table 2, the ROC AUC values increased from 0.821 to 0.932 and ...

  2. [10]

    Results Validated with Clinician-Generated Labels Assessment of Predicted and Manual Label Uncertainty As detailed in Section II.A.2b, we grouped the test data by the uncertainty derived from manual labels from three clinicians and compared these values to the uncertainty esti...

  3. [11]

    The results indicate an initially positive trend, with accuracy increasing as the sample size grows from 10 to 30

    Transfer Learning with Limited Manual Labels Table 6 illustrates the influence of training sample size on the overall accuracy of transfer learning from our BOC model pre-trained using surrogate labels. The results indicate an initially positive trend, with accuracy increasing...

  4. [12]

    This case exhibits high uncertainty for the AI model; please review carefully

    Training with Manual Labels from Scratch Increasing the sample size clearly improves our model’s performance when trained from scratch using manual labels. Specifically, expanding the training sample size to 80 subjects increased the model's accuracy to 85.6%, though it remain...

  5. [13]

    Exploring the cancer patients’ experiences during external radiotherap y: A sys- tematic review and thematic synthesis of qualitative and quantitative evidence,

    A. M. Grilo, B. Santos, I. Baptista, and F. Monsanto, “Exploring the cancer patients’ experiences during external radiotherap y: A sys- tematic review and thematic synthesis of qualitative and quantitative evidence,” Eur. J. Oncol. Nurs., vol. 52, p. 101965, Jun. 2021, doi: 10...

  6. [14]

    Online adaptive radiotherapy potentially reduces toxicity for high-risk prostate cancer treatment,

    R. L. Christiansen et al., “Online adaptive radiotherapy potentially reduces toxicity for high-risk prostate cancer treatment,” Radiother. Oncol., vol. 167, pp. 165–171, Feb. 2022, doi: 10.1016/j.radonc.2021.12.013

  7. [16]

    Metrics to evaluate the performance of auto -segmentation for radiation treatment planning: A critical review,

    M. V. Sherer et al., “Metrics to evaluate the performance of auto -segmentation for radiation treatment planning: A critical review,” Radiother. Oncol., vol. 160, pp. 185–191, Jul. 2021, doi: 10.1016/j.radonc.2021.05.003

  8. [17]

    A framework for automated contour quality assurance in radiation therapy including adaptive techniques,

    M. B. Altman et al., “A framework for automated contour quality assurance in radiation therapy including adaptive techniques,” Phys. Med. Biol., vol. 60, no. 13, pp. 5199–5209, Jul. 2015, doi: 10.1088/0031-9155/60/13/5199

  9. [18]

    Contouring quality assurance methodology based on multiple geometric features against deep learning auto‐segmenta- tion,

    J. Duan et al., “Contouring quality assurance methodology based on multiple geometric features against deep learning auto‐segmenta- tion,” Med. Phys., vol. 50, no. 5, pp. 2715–2732, May 2023, doi: 10.1002/mp.16299

  10. [19]

    Comprehensive Clinical Usability-Oriented Contour Quality Evaluation for Deep Learning Auto-segmentation: Com- bining Multiple Quantitative Metrics Through Machine Learning,

    Y. Zhang et al., “Comprehensive Clinical Usability-Oriented Contour Quality Evaluation for Deep Learning Auto-segmentation: Com- bining Multiple Quantitative Metrics Through Machine Learning,” Pract. Radiat. Oncol., vol. 15, no. 1, pp. 93 –102, Jan. 2025, doi: 10.1016/j.prro.2...

  11. [20]

    Automated Quality Assurance of OAR Contouring for Lung Cancer Based on Segmentation With Deep Active Learning,

    K. Men, H. Geng, T. Biswas, Z. Liao, and Y. Xiao, “Automated Quality Assurance of OAR Contouring for Lung Cancer Based on Segmentation With Deep Active Learning,” Front. Oncol., vol. 10, p. 986, Jul. 2020, doi: 10.3389/fonc.2020.00986

  12. [21]

    CNN-Based Quality Assurance for Automatic Segmentation of Breast Cancer in Radiotherapy,

    X. Chen et al., “CNN-Based Quality Assurance for Automatic Segmentation of Breast Cancer in Radiotherapy,” Front. Oncol., vol. 10, p. 524, Apr. 2020, doi: 10.3389/fonc.2020.00524

  13. [22]

    Uncertainty Assessment for Deep Learning Radiotherapy Applications,

    C. A. T. Van Den Berg and E. F. Meliadò, “Uncertainty Assessment for Deep Learning Radiotherapy Applications,” Semin. Radiat. Oncol., vol. 32, no. 4, pp. 304–318, Oct. 2022, doi: 10.1016/j.semradonc.2022.06.001

  14. [23]

    Towards reliable head and neck cancers locoregional recurrence prediction using delta-radiomics and learning with rejection option,

    K. Wang, M. Dohopolski, Q. Zhang, D. Sher, and J. Wang, “Towards reliable head and neck cancers locoregional recurrence prediction using delta-radiomics and learning with rejection option,” Med. Phys., vol. 50, no. 4, pp. 2212–2223, 2023, doi: 10.1002/mp.16132

  15. [24]

    Predicting lymph node metastasis in patients with oropharyngeal cancer by using a convolutional neural network with associated epistemic and aleatoric uncertainty,

    M. Dohopolski, L. Chen, D. S her, and J. Wang, “Predicting lymph node metastasis in patients with oropharyngeal cancer by using a convolutional neural network with associated epistemic and aleatoric uncertainty,” Phys. Med. Biol., vol. 65, no. 22, p. 225002, Nov. 2020, doi: 10...

  16. [25]

    Uncertainty estimations methods for a deep learning model to aid in clinical decision -making – a clinician’s perspective

    M. Dohopolski et al., “Uncertainty estimations methods for a deep learning model to aid in clinical decision -making – a clinician’s perspective”

  17. [26]

    A deep learning-based framework for segmenting invisible clinical target volumes with estimated uncertainties for post-operative prostate cancer radiotherapy,

    A. Balagopal et al., “A deep learning-based framework for segmenting invisible clinical target volumes with estimated uncertainties for post-operative prostate cancer radiotherapy,” Med. Image Anal., vol. 72, p. 102101, Aug. 2021, doi: 10.1016/j.media.2021.102101

  18. [27]

    Uncertainty quantification using Bayesian neural networks in classification: Applica- tion to biomedical image segmentation,

    Y. Kwon, J.-H. Won, B. J. Kim, and M. C. Paik, “Uncertainty quantification using Bayesian neural networks in classification: Applica- tion to biomedical image segmentation,” Comput. Stat. Data Anal., vol. 142, p. 106816, Feb. 2020, doi: 10.1016/j.csda.2019.106816

  19. [28]

    Uncertainty-driven Sanity Check: Application to Postoperative Brain Tumor Cavity Segmentation,

    A. Jungo, R. Meier, E. Ermis, E. Herrmann, and M. Reyes, “Uncertainty-driven Sanity Check: Application to Postoperative Brain Tumor Cavity Segmentation,” Jun. 08, 2018, arXiv: arXiv:1806.03106. Accessed: Apr. 17, 2024. [Online]. Available: http://arxiv.org/abs/1806.03106

  20. [29]

    Using Spatial Probability Maps to Hi ghlight Potential Inaccuracies in Deep Learning -Based Contours: Facilitating Online Adaptive Radiation Therapy - ClinicalKey

    “Using Spatial Probability Maps to Hi ghlight Potential Inaccuracies in Deep Learning -Based Contours: Facilitating Online Adaptive Radiation Therapy - ClinicalKey.” Accessed: Apr. 17, 2024. [Online]. Available: https://www.clinicalkey.com/#!/content/playCon- tent/1-s2.0-S2452...

  21. [30]

    A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty estimation in radiation therapy dose prediction with deep learning neural networks,

    D. Nguyen et al., “A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty estimation in radiation therapy dose prediction with deep learning neural networks,” Phys. Med. Biol. , vol. 66, no. 5, p. 054002, Feb. 2021, doi: 10.1088/1361-6...

  22. [31]

    Uncertainty in Multitask Learning: Joint Representations for Probabilistic MR -only Radiotherapy Planning,

    F. J. S. Bragman et al., “Uncertainty in Multitask Learning: Joint Representations for Probabilistic MR -only Radiotherapy Planning,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, A. F. Frangi, J. A. Schnabel, C. Davatzikos, C. Alberola-López, an...

  23. [32]

    Deep learning based synthetic-CT generation in radiotherapy and PET: A review,

    M. F. Spadea, M. Maspero, P. Zaffino, and J. Seco, “Deep learning based synthetic-CT generation in radiotherapy and PET: A review,” Med. Phys., vol. 48, no. 11, pp. 6537–6566, 2021, doi: 10.1002/mp.15150

  24. [33]

    Deep Generative Model for Synthetic-CT Generation with Uncertainty Predictions,

    M. Hemsley et al., “Deep Generative Model for Synthetic-CT Generation with Uncertainty Predictions,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2020, A. L. Martel, P. Abolmaesumi, D. Stoyanov, D. Mateus, M. A. Zuluaga, S. K. Zhou, D. Racoceanu, a nd...

  25. [34]

    Uncertainty in Deep Learning,

    Y. Gal, “Uncertainty in Deep Learning,” University of Cambridge. [Online]. Available: chrome -extension://efaidnbmnnnibpcaj- pcglclefindmkaj/https://www.cs.ox.ac.uk/people/yarin.gal/website/thesis/thesis.pdf

  26. [35]

    Deep neural networks for rank -consistent ordinal regression based on conditional probabilities,

    X. Shi, W. Cao, and S. Raschka, “Deep neural networks for rank -consistent ordinal regression based on conditional probabilities,” Pattern Anal. Appl., vol. 26, no. 3, pp. 941–955, Aug. 2023, doi: 10.1007/s10044-023-01181-9

  27. [36]

    Clinical evaluation of a deep learning segmentation model including manual adjustments afterwards for locally advanced breast cancer,

    N. Bakx et al., “Clinical evaluation of a deep learning segmentation model including manual adjustments afterwards for locally advanced breast cancer,” Tech. Innov. Patient Support Radiat. Oncol., vol. 26, p. 100211, May 2023, doi: 10.1016/j.tipsro.2023.100211

  28. [37]

    Automated Contouring and Planning in Radiation Therapy: What Is ‘Clinically Acceptable’?,

    H. Baroudi et al., “Automated Contouring and Planning in Radiation Therapy: What Is ‘Clinically Acceptable’?,” Diagnostics, vol. 13, no. 4, p. 667, Feb. 2023, doi: 10.3390/diagnostics13040667

  29. [38]

    Incremental retraining, clinical implementation, and acceptance rate of deep learning auto‐segmentation for male pelvis in a multiuser environment,

    J. Duan et al., “Incremental retraining, clinical implementation, and acceptance rate of deep learning auto‐segmentation for male pelvis in a multiuser environment,” Med. Phys., vol. 50, no. 7, pp. 4079–4091, Jul. 2023, doi: 10.1002/mp.16537

  30. [39]

    Evaluating the clinical acceptability of deep learning contours of prostate and organs -at-risk in an automated prostate treatment planning process,

    J. Duan et al., “Evaluating the clinical acceptability of deep learning contours of prostate and organs -at-risk in an automated prostate treatment planning process,” Med. Phys., vol. 49, no. 4, pp. 2570–2581, 2022, doi: 10.1002/mp.15525

  31. [40]

    Clinical implementation of deep learning contour autosegmentation for prostate radiotherapy,

    E. Cha et al., “Clinical implementation of deep learning contour autosegmentation for prostate radiotherapy,” Radiother. Oncol., vol. 159, pp. 1–7, Jun. 2021, doi: 10.1016/j.radonc.2021.02.040

  32. [41]

    Deep learning to achieve clinically applicable segmenta tion of head and neck anatomy for radiotherapy,

    S. Nikolov et al., “Deep learning to achieve clinically applicable segmenta tion of head and neck anatomy for radiotherapy,” Jan. 13, 2021, arXiv: arXiv:1809.04430. Accessed: Apr. 11, 2024. [Online]. Available: http://arxiv.org/abs/1809.04430

  33. [42]

    A systematic review of deep learning data augmentation in medical imaging: Recent advances and future research directions,

    T. Islam, Md. S. Hafiz, J. R. Jim, Md. M. Kabir, and M. F. Mridha, “A systematic review of deep learning data augmentation in medical imaging: Recent advances and future research directions,” Healthc. Anal. , vol. 5, p. 100340, Jun. 2024, doi: 10.1016/j.health.2024.100340

  34. [43]

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning,

    Y. Gal and Z. Ghahramani, “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning,” in Proceedings of The 33rd International Conference on Machine Learning , PMLR, Jun. 2016, pp. 1050 –1059. Accessed: Feb. 23, 2025. [Online]. Available: https://pr...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.