Pith. sign in

REVIEW 1 major objections 1 minor 50 references

Comparative Analysis of Pre-trained Deep Learning Models and DINOv2 for Cushing's Syndrome Diagnosis in Facial Analysis

T0 review · 1 major / 1 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read Transformer-based vision models and the DINOv2 foundation model outperform CNNs in diagnosing Cushing's syndrome from facial photographs, with ViT reaching the top F1 score.

desk verdict A useful pilot benchmark for Cushing's facial diagnosis, but the ViT-vs-DINOv2-vs-CNN ranking rests on 16 positive test images and a single split, so the headline ordering is statistically fragile. read the letter →

arxiv 2501.12023 v1 pith:CNI3SZH7 submitted 2025-01-21 cs.LG cs.CVeess.IV

classification cs.LGcs.CVeess.IV
keywords Cushing'ssyndromefacialimageclassificationvisiontransformerDINOv2transferlearningparameterfreezinggenderbiasfoundationmodels
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 sets out to establish that Transformer-based vision models and the DINOv2 foundation model are better starting points than convolutional neural networks for diagnosing Cushing's syndrome from frontal facial photographs. On a single-center clinical dataset of 343 subjects (49 patients, 294 controls), the authors report ViT as the best classifier, with an F1 score of 85.74% (85.71% in the results table), and report that every frozen DINOv2 variant reached at least 80% F1, while the best CNN reached 71.11%. They also report that freezing DINOv2's backbone raised its F1 by 4.93 to 18.46 percentage points and that every model scored higher on female than on male faces. If these results hold, attention-based architectures and frozen self-supervised features are a more sensible default than CNNs for facial screening of endocrine disorders on small clinical datasets, and the reported gender gap would need to be addressed before clinical deployment.

What carries the argument

The central objects being compared are the architectures themselves: convolutional networks (DenseNet, ResNet) that read images through local receptive fields, Transformer-based classifiers (ViT, Swin) whose self-attention mechanism relates every part of the face to every other part, and DINOv2, a self-supervised foundation model built on the Vision Transformer and trained without labels by self-distillation. The carrying mechanism of the study is transfer learning: every model starts from ImageNet pre-training and is then fine-tuned on the Cushing's dataset, while DINOv2 is additionally tested with its backbone frozen and only the classifier head trained. A preprocessing pipeline using dlib face detection and SAM segmentation strips the background before images are resized to 224x224, and the disease class is augmented fivefold in the training set to counter the 1:6 class imbalance.

What would settle it

Collect a multi-center test set of frontal facial photos taken with a standardized camera and lighting protocol, with at least 30 male disease patients, and retrain the same models on one center while testing on another; if ViT no longer beats the CNNs, or if the gender gap disappears once male training data are balanced, the paper's central and gender claims are contradicted.

Watch

Extended reading notes

Core claim

The central claim is that in transfer learning for Cushing's syndrome facial diagnosis, the inductive bias of convolutional neural networks toward local features is a disadvantage because the disease presents as global facial changes such as moon facies and plethora, whereas self-attention models such as ViT and Swin, and the DINOv2 foundation model, capture long-range dependencies and therefore classify better. The authors report the highest F1 score for ViT-B-32 at 85.74% in the abstract (85.71% in Table II), with frozen DINOv2-b achieving 82.35%, and all Transformer and foundation models exceeding 80% F1 while best DenseNet reached 70.00% and best ResNet 71.11%. They further claim that parameter freezing improves DINOv2 performance, with F1 gains of 16.84, 4.93, and 18.46 percentage points across the small, base, and large variants, and that all models exhibit gender bias traceable to the female-majority training set. Activation-map inspection is used to argue that the better models focus on facial regions consistent with medical knowledge of Cushing's syndrome.

Load-bearing premise

The load-bearing premise is that the only systematic difference between disease and control photographs is the disease itself, not differences in camera, lighting, age, or obesity-related facial fullness, and that the four male disease patients in the test set are enough to measure how well the models work for men.

Editorial extensions

If this is right

  • Clinical facial screening studies for endocrine disorders should benchmark Transformer and foundation models instead of assuming CNN features are sufficient, because global facial changes are the disease's visible signature.
  • Freezing DINOv2's backbone is a practical training choice: the paper reports higher F1 scores and faster training than unfrozen fine-tuning on a small dataset.
  • Male patients need explicit data collection before any of these models is used clinically, since every model performed markedly worse on male faces.
  • Activation maps of the better models concentrate on disease-relevant facial regions, so the classifiers could double as a visual aid for clinicians rather than only a binary decision.
  • Bigger is not automatically better: the best performer in each family was a mid-size variant (ViT-B-32, DINOv2-b, ResNet-34, Swin-B), so model selection should be empirical.

Reading between the lines

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

  • Beyond the reported experiments, an implicit testable extension is whether the ViT/DINOv2 advantage persists across hospitals and camera protocols, or whether some of the gap is site-specific rather than disease-specific.
  • The freezing result suggests a cheaper general recipe for small clinical datasets, but the paper does not test whether other self-supervised backbones also benefit from freezing; that comparison would separate the architecture effect from the DINOv2-specific effect.
  • Because the gender result is computed from only four male disease patients in the test set, balancing male training data or collecting more male patients could either confirm or overturn the claimed bias.
  • An untested alternative is whether data augmentation with synthetically generated disease faces, rather than fivefold repetition of the same 49 images, would close the gap between CNNs and Transformers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 1 minor

Summary. The paper presents an empirical comparison of pre-trained deep learning models—DenseNet, ResNet, ViT, Swin Transformer, and DINOv2—for diagnosing Cushing's syndrome from frontal facial images. Using a single-center dataset of 343 images (49 disease, 294 control) split into training, validation, and test sets at a 1:1:1 ratio, the authors fine-tune each model and report accuracy, precision, recall, and F1. They report that Transformer-based models and DINOv2 outperform CNNs, with ViT-B32 achieving the highest F1 of 85.71%, that freezing DINOv2's backbone improves its F1, and that model performance is higher for female than for male samples. The authors also show activation maps to argue that ViT and DINOv2 focus on medically relevant facial regions. The code is publicly available.

Significance. If the findings were statistically robust, this study would provide a useful benchmark for model selection in facial-image screening of Cushing's syndrome, a rare disease where data are scarce. The comparison covers a broad set of architectures, and the public release of code is a practical strength. The paper also addresses an understudied question: whether vision foundation models (DINOv2) offer an advantage over ImageNet-pretrained CNNs in small clinical datasets. However, the significance is currently limited by the very small test set (16 disease positives), the absence of confidence intervals or significance tests, and a single random split, which make the headline ordering of models and the gender-bias conclusion fragile.

major comments (1)
  1. [Section IV-A and Section V] The authors interpret the models' high accuracy as evidence that the models learn the Cushing's facial phenotype, and they use activation maps in Section V to argue that the models focus on "key areas in facial images that align with medical knowledge." This interpretation presupposes that the control and disease groups differ only in disease status. Section IV-A states that controls were obese, overweight, or BMI-matched and had no typical Cushing's features, but no demographic balance table beyond gender, no acquisition-protocol standardization, and no control for age, camera, lighting, or residual BMI differences are reported. If the groups differ systematically in these non-disease attributes, the classifier could be learning confounds. The activation-map analysis is post hoc interpretation, not a prediction from the model, and does not rule out confounding.
minor comments (1)
  1. [Section V] The manuscript states that "all four ViT models (ViT-B-16, ViT-B-32, ViT-L-16) exceeded an F1 score of 80%," but the list names only three models; ViT-L-32 has F1=0.7407 in Table II, which contradicts the statement as written.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark with external pretrained models, held-out test set, and no fitted parameter renamed as a prediction.

full rationale

This paper is a comparative empirical evaluation of pretrained CNN, ViT, Swin, and DINOv2 models on a fixed clinical dataset of facial images for Cushing's syndrome classification. The central claims—that Transformer-based models and DINOv2 outperform CNNs, that ViT-B32 has the highest F1 score (85.71%), and that freezing DINOv2's backbone improves performance—are all computed from held-out test-set predictions reported in Tables II and III, using external pretrained checkpoints and standard fine-tuning. There is no derivation chain in which an output is constructed from its own inputs: no parameter is fitted to the test set and then reported as a prediction; the activation-map commentary in Section V is post hoc interpretation rather than a prediction derived from fitted quantities; and no load-bearing uniqueness theorem or author-imported ansatz is invoked. The self-citations in the reference list are ordinary citations to prior work and do not carry the argument. The paper's limitations—small sample size, single-center data, gender imbalance, and potential confounding between groups—are acknowledged in Section VI and weaken the statistical strength of the empirical comparison, but they are not circularity. The appropriate circularity score is therefore 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The main load-bearing inputs are the clinical dataset's labeling and matching assumptions, plus the hand-chosen training hyperparameters. There is no fitted equation or free constant in the traditional sense, but the fixed training protocol is a set of free choices that the reported performance depends on.

free parameters (5)
  • learning rate = 5e-4
    Fixed for all models; no per-model tuning or schedule reported, so it is a hand-chosen training hyperparameter that affects all reported F1 scores.
  • batch size = 32
    Hand-chosen for all models; affects optimization and final performance.
  • training epochs = 200
    Fixed stopping point; early stopping is not reported, so results depend on this choice.
  • disease augmentation multiplier = 5x (17 to 85 training samples)
    Disease-class training images were augmented to five times the original count to approximate class balance; the multiplier is chosen by hand and shapes the learned decision boundary.
  • input image size = 224x224
    All face images were resampled to this fixed resolution; the choice affects the feature detail available to the models.
assumptions (5)
  • domain assumption Frontal facial photographs contain sufficient diagnostic signal to classify Cushing's syndrome.
    The entire study presupposes that 2D frontal images alone are enough for automated diagnosis, following prior work by Popp et al. and Wei et al.
  • domain assumption The control group is matched to the disease group on BMI and differs only in the absence of Cushing's features.
    Section IV-A states that controls were obese, overweight, or BMI-matched and had no typical Cushing's features; if matching is imperfect, the classifier could learn confounds rather than disease phenotype.
  • domain assumption Dlib plus SAM face extraction preserves all disease-relevant facial features while removing background.
    Section IV-C describes cropping and segmentation; if the pipeline discards periorbital or perioral cues, model comparisons are affected.
  • domain assumption ImageNet and self-supervised pretrained features transfer to clinical facial photographs.
    All models are used as pretrained backbones without domain-specific pretraining; transferability is assumed.
  • domain assumption The random 1:1:1 split produces an independent and representative test set.
    Section IV-A describes a single random split; no stratification by gender or age is reported, and the male disease test subset has 4 samples.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Comparative Analysis of Pre-trained Deep Learning Models and DINOv2 for Cushing's Syndrome Diagnosis in Facial Analysis." pith.science (2026). https://pith.science/paper/CNI3SZH7

@misc{pith2026250112023,
  author       = {Pith},
  title        = {Pith review of: Comparative Analysis of Pre-trained Deep Learning Models and DINOv2 for Cushing's Syndrome Diagnosis in Facial Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CNI3SZH7}},
  note         = {Machine review of arXiv:2501.12023}
}
read the original abstract

Cushing's syndrome is a condition caused by excessive glucocorticoid secretion from the adrenal cortex, often manifesting with moon facies and plethora, making facial data crucial for diagnosis. Previous studies have used pre-trained convolutional neural networks (CNNs) for diagnosing Cushing's syndrome using frontal facial images. However, CNNs are better at capturing local features, while Cushing's syndrome often presents with global facial features. Transformer-based models like ViT and SWIN, which utilize self-attention mechanisms, can better capture long-range dependencies and global features. Recently, DINOv2, a foundation model based on visual Transformers, has gained interest. This study compares the performance of various pre-trained models, including CNNs, Transformer-based models, and DINOv2, in diagnosing Cushing's syndrome. We also analyze gender bias and the impact of freezing mechanisms on DINOv2. Our results show that Transformer-based models and DINOv2 outperformed CNNs, with ViT achieving the highest F1 score of 85.74%. Both the pre-trained model and DINOv2 had higher accuracy for female samples. DINOv2 also showed improved performance when freezing parameters. In conclusion, Transformer-based models and DINOv2 are effective for Cushing's syndrome classification.

Figures

Figures reproduced from arXiv: 2501.12023 by the authors.

Figure 1
Figure 1. Flow diagram of this study. DINOv2, BLIP [34], and CLIP [35]—across four comprehen￾sive medical imaging datasets. When paired with classifiers, DINOv2 outperformed the baseline, demonstrating strong ap￾plicability in medical imaging. These studies underscore the potential and effectiveness of the DINOv2 model, emphasizing its significant role in medical image analysis. However, it is important to note that most of t… view at source ↗
Figure 2
Figure 2. Activation maps of the pre-trained model and the foundational model. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 18 canonical work pages

  1. [1]

    Diagnosis and complications of cushing’s syndrome: a consensus state- ment,

    G. Arnaldi, A. Angeli, A. Atkinson, X. Bertagna, F. Cavagnini, G. Chrousos, G. Fava, J. Findling, R. Gaillard, A. Grossman et al. , “Diagnosis and complications of cushing’s syndrome: a consensus state- ment,” The Journal of Clinical Endocrinology & Metabolism , vol. 88, no. 12, pp. 5593–5602, 2003

  2. [2]

    Persistence of myopathy in cushing’s syndrome: evaluation of the german cushing’s registry,

    C. M. Berr, M. R. Stieg, T. Deutschbein, M. Quinkler, R. Schmidmaier, A. Osswald, N. Reisch, K. Ritzel, C. Dimopoulou, J. Fazel et al. , “Persistence of myopathy in cushing’s syndrome: evaluation of the german cushing’s registry,”European journal of endocrinology, vol. 176, no. 6, pp. 737–746, 2017

  3. [3]

    Demographic factors and the presence of comorbidities do not promote early detection of cushing’s disease and acromegaly,

    T. Psaras, M. Milian, V . Hattermann, T. Freiman, B. Gallwitz, and J. Honegger, “Demographic factors and the presence of comorbidities do not promote early detection of cushing’s disease and acromegaly,” Experimental and clinical endocrinology & diabetes , vol. 119, no. 01, pp. 21–25, 2011

  4. [4]

    Computer vision technology in the differential diagnosis of cushing’s syndrome,

    K. H. Popp, R. P. Kosilek, R. Frohner, G. K. Stalla, A. Athanasoulia- Kaspar, C. Berr, S. Zopp, M. Reincke, M. Witt, R. P. W ¨urtz et al. , “Computer vision technology in the differential diagnosis of cushing’s syndrome,” Experimental and Clinical Endocrinology & Diabetes , vol. 127, no. 10, pp. 685–690, 2019

  5. [5]

    Deep-learning approach to automatic identifi- cation of facial anomalies in endocrine disorders,

    R. Wei, C. Jiang, J. Gao, P. Xu, D. Zhang, Z. Sun, X. Liu, K. Deng, X. Bao, G. Sun et al. , “Deep-learning approach to automatic identifi- cation of facial anomalies in endocrine disorders,” Neuroendocrinology, vol. 110, no. 5, pp. 328–337, 2020

  6. [6]

    Dlib-ml: A machine learning toolkit,

    D. E. King, “Dlib-ml: A machine learning toolkit,” The Journal of Machine Learning Research , vol. 10, pp. 1755–1758, 2009

  7. [7]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

  8. [8]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

Show all 50 references
  1. [9]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

  2. [10]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

  3. [11]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint arXiv:1907.11692, 2019

  4. [12]

    Improving language understanding by generative pre- training,

    A. Radford, “Improving language understanding by generative pre- training,” 2018

  5. [13]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

  6. [14]

    Language models are few-shot learners,

    T. B. Brown, “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165, 2020

  7. [16]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022. 10

  8. [17]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al. , “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

  9. [18]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

  10. [19]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,” arXiv preprint arXiv:...

  11. [20]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,” Nature Communications, vol. 15, pp. 1–9, 2024

  12. [21]

    Are natural domain foundation models useful for medical image classification?

    J. P. Huix, A. R. Ganeshan, J. F. Haslum, M. S ¨oderberg, C. Matsoukas, and K. Smith, “Are natural domain foundation models useful for medical image classification?” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 7634–7643

  13. [22]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE confer- ence on computer vision and pattern recognition , 2017, pp. 4700–4708

  14. [23]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

  15. [24]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    D. Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv: 2010.11929 , 2020

  16. [25]

    Medical sam adapter: Adapting segment anything model for medical image segmentation,

    J. Wu, W. Ji, Y . Liu, H. Fu, M. Xu, Y . Xu, and Y . Jin, “Medical sam adapter: Adapting segment anything model for medical image segmentation,” arXiv preprint arXiv:2304.12620 , 2023

  17. [26]

    Ma-sam: Modality-agnostic sam adaptation for 3d medical image segmentation,

    C. Chen, J. Miao, D. Wu, A. Zhong, Z. Yan, S. Kim, J. Hu, Z. Liu, L. Sun, X. Li et al. , “Ma-sam: Modality-agnostic sam adaptation for 3d medical image segmentation,” Medical Image Analysis , p. 103310, 2024

  18. [27]

    Sam-med2d,

    J. Cheng, J. Ye, Z. Deng, J. Chen, T. Li, H. Wang, Y . Su, Z. Huang, J. Chen, L. Jiang et al., “Sam-med2d,” arXiv preprint arXiv:2308.16184, 2023

  19. [28]

    Input augmentation with sam: Boosting medical image segmentation with seg- mentation foundation model,

    Y . Zhang, T. Zhou, S. Wang, P. Liang, Y . Zhang, and D. Z. Chen, “Input augmentation with sam: Boosting medical image segmentation with seg- mentation foundation model,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023,...

  20. [29]

    Vision-mae: A foundation model for medical image seg- mentation and classification,

    Z. Liu, A. Tieu, N. Patel, A. Zhou, G. Soultanidis, Z. A. Fayad, T. Deyer, and X. Mei, “Vision-mae: A foundation model for medical image seg- mentation and classification,” arXiv preprint arXiv:2402.01034 , 2024

  21. [30]

    Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,

    M. Baharoon, W. Qureshi, J. Ouyang, Y . Xu, K. Phol, A. Aljouie, and W. Peng, “Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,” arXiv preprint arXiv:2312.02366 , 2023

  22. [31]

    Parameter-efficient fine-tuning of dinov2 vision transformers for lung nodule classification,

    B. P. Veasey and A. A. Amini, “Parameter-efficient fine-tuning of dinov2 vision transformers for lung nodule classification,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) . IEEE, 2024, pp. 1–5

  23. [32]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

  24. [33]

    Segment everything everywhere all at once,

    X. Zou, J. Yang, H. Zhang, F. Li, L. Li, J. Wang, L. Wang, J. Gao, and Y . J. Lee, “Segment everything everywhere all at once,” Advances in Neural Information Processing Systems , vol. 36, 2024

  25. [34]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” in International conference on machine learning . PMLR, 2022, pp. 12 888–12 900

  26. [35]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

  27. [36]

    Electronic medical records as input to predict postoperative immediate remission of cushing’s disease: application of word embedding,

    W. Zhang, D. Li, M. Feng, B. Hu, Y . Fan, Q. Chen, and R. Wang, “Electronic medical records as input to predict postoperative immediate remission of cushing’s disease: application of word embedding,” Fron- tiers in Oncology , vol. 11, p. 754882, 2021

  28. [37]

    Learning repre- sentations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- sentations by back-propagating errors,” nature, vol. 323, no. 6088, pp. 533–536, 1986

  29. [38]

    Support-vector networks,

    C. Cortes, “Support-vector networks,” Machine Learning, 1995

  30. [39]

    Random forests,

    L. Breiman, “Random forests,” Machine learning , vol. 45, pp. 5–32, 2001

  31. [40]

    The regression analysis of binary sequences,

    D. R. Cox, “The regression analysis of binary sequences,” Journal of the Royal Statistical Society Series B: Statistical Methodology , vol. 20, no. 2, pp. 215–232, 1958

  32. [41]

    Toward better prediction of recurrence for cushing’s disease: a factorization-machine based neural approach,

    Y . Fan, D. Li, Y . Liu, M. Feng, Q. Chen, and R. Wang, “Toward better prediction of recurrence for cushing’s disease: a factorization-machine based neural approach,” International Journal of Machine Learning and Cybernetics, vol. 12, pp. 625–633, 2021

  33. [42]

    Greedy function approximation: a gradient boosting machine,

    J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of statistics , pp. 1189–1232, 2001

  34. [43]

    A decision-theoretic generalization of on-line learning and an application to boosting,

    Y . Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning and an application to boosting,” Journal of computer and system sciences , vol. 55, no. 1, pp. 119–139, 1997

  35. [44]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , 2016, pp. 785–794

  36. [45]

    Machine learn- ing models for classification of cushing’s syndrome using retrospective data,

    S. Isci, D. S. Y . Kalender, F. Bayraktar, and A. Yaman, “Machine learn- ing models for classification of cushing’s syndrome using retrospective data,” IEEE Journal of Biomedical and Health Informatics , vol. 25, no. 8, pp. 3153–3162, 2021

  37. [46]

    An introduction to kernel and nearest-neighbor non- parametric regression,

    N. S. Altman, “An introduction to kernel and nearest-neighbor non- parametric regression,” The American Statistician , vol. 46, no. 3, pp. 175–185, 1992

  38. [47]

    Principal component analysis-a tutorial,

    A. Tharwat, “Principal component analysis-a tutorial,” International Journal of Applied Pattern Recognition , vol. 3, no. 3, pp. 197–240, 2016

  39. [48]

    Classification and regression trees,

    W.-Y . Loh, “Classification and regression trees,” Wiley interdisciplinary reviews: data mining and knowledge discovery , vol. 1, no. 1, pp. 14–23, 2011

  40. [49]

    Automatic face classification of cushing’s syndrome in women–a novel screening approach,

    R. Kosilek, J. Schopohl, M. Grunke, M. Reincke, C. Dimopoulou, G. Stalla, R. W ¨urtz, A. Lammert, M. G ¨unther, and H. Schneider, “Automatic face classification of cushing’s syndrome in women–a novel screening approach,” Experimental and Clinical Endocrinology & Dia- betes, vo...

  41. [50]

    Large-scale machine learning with stochastic gradient de- scent,

    L. Bottou, “Large-scale machine learning with stochastic gradient de- scent,” in Proceedings of COMPSTAT’2010: 19th International Con- ference on Computational StatisticsParis France, August 22-27, 2010 Keynote, Invited and Contributed Papers . Springer, 2010, pp. 177– 186

  42. [51]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.