Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Hybrid Ensemble Approaches: Optimal Deep Feature Fusion and Hyperparameter-Tuned Classifier Ensembling for Enhanced Brain Tumor Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a double ensemble—fused deep features plus voted tuned classifiers—reaches up to 99.4% accuracy on brain-tumor MRI datasets.

desk verdict Broad, careful benchmark undermined by test-set selection: the reported SOTA accuracies don't estimate generalization. read the letter →

arxiv 2507.12177 v1 pith:E2ZCUI2Q submitted 2025-07-16 cs.CV

classification cs.CV
keywords braintumorclassificationMRItransferlearningfeatureensembleclassifierhyperparametertuningvisiontransformer
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

The paper sets out to show that brain-tumor classification from MRI improves when two ensemble layers are stacked: a feature layer that concatenates deep features from the top two or three pre-trained CNN and vision-transformer models, and a classifier layer that pools the predictions of top-performing tuned machine-learning classifiers. The authors report that this double ensembling, after grid-search hyperparameter tuning, reaches accuracies of 97.5% on a small binary dataset, 99.43% on a large binary dataset, and 99.33% on a large four-class dataset, surpassing the individual models and the earlier results they compare against. This matters because, if correct, relatively lightweight tuned classifiers on frozen pretrained features can rival end-to-end deep networks on small medical image datasets, which is the common practical setting in medical imaging.

What carries the argument

The machinery is a two-stage ensemble. Stage one selects the best frozen feature extractors by average accuracy over nine tuned ML classifiers, then concatenates the top two or three feature vectors into a single fused vector. Stage two trains the nine classifiers on that fused vector and, for the final prediction, pools the top two or three classifiers by voting; the enhanced variant also applies Min-Max normalization, PCA retaining half of the components, and SMOTE oversampling before classification.

What would settle it

Split each dataset once into training, validation, and test; choose top feature extractors and classifier ensembles on validation only; then measure test accuracy. If the double-ensemble test accuracies fall materially below 97.5%, 99.43%, and 99.33%, the headline numbers partly reflect selection on the test set.

Watch

Extended reading notes

Core claim

The central claim is that feature-level ensembling and classifier-level ensembling are complementary, and that tuning the classifiers' hyperparameters is what unlocks most of the gain. After cropping, resizing, and augmentation, twenty-five frozen ImageNet models—twelve CNNs and thirteen vision transformers—produce deep feature vectors; nine machine-learning classifiers with grid-searched hyperparameters score each feature set, and the top two or three models' features are concatenated. The concatenated features are then classified by single tuned classifiers or by voted ensembles of the top two or three classifiers, with normalization, PCA, and SMOTE applied as enhanced variants. Across three public MRI datasets, the best combinations are a tuned SVM-RBF plus MLP or SVM-linear vote on the small binary set, a KNN plus SVM-RBF plus MLP vote on the large binary set, and an RF plus MLP plus KNN vote on the four-class set, yielding the highest accuracies the paper reports for each dataset.

Load-bearing premise

The load-bearing premise is that the top feature extractors and the top classifiers were chosen without ever looking at the test set, even though the paper's selection step reports accuracies on that same test split; if test information leaked into selection, the reported gains are optimistic.

Editorial extensions

If this is right

  • Tuning each classifier with grid search raises accuracy more than ensembling alone, so tuning before combining is a practical baseline.
  • Fusing features from different pretrained architectures produces a discriminative vector that small datasets can use without training a deep network from scratch.
  • The best voted classifier ensemble differs by task: SVM-family classifiers for binary cases and RF plus MLP plus KNN for the four-class case.
  • The pipeline transfers to other small medical image classification tasks where pretrained feature extractors and tuned classifiers are available.

Reading between the lines

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

  • Because the top feature extractors and classifiers are selected on the same test split that reports final accuracy, a fair reader should expect the headline numbers to shrink under nested validation or a truly held-out test set.
  • The paper does not test cross-dataset generalization; one natural extension is to train the selection pipeline on one MRI dataset and evaluate on another, directly testing whether the fused features transfer.
  • A cleaner isolation of the ensembling contribution would compare the fused-feature ensemble against the best single feature extractor with the same tuned classifier, using a validation-based selection rather than the same-split accuracy.
  • The four-class gains are concentrated in the RF, MLP, and KNN vote; testing whether this vote also helps on larger multi-class medical datasets would show whether the pattern is task-specific or general.
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

3 major / 5 minor

Summary. The paper proposes a two-level ensemble framework for brain tumor classification from MRI: 25 pre-trained CNN and ViT models are used as frozen feature extractors, nine ML classifiers are trained on the resulting deep features, and the top-2/top-3 feature extractors and top-2/top-3 classifiers are ensembled to improve accuracy. Hyperparameters are tuned by grid search, and four ablations (raw features, normalization+PCA, SMOTE, and normalization+PCA+SMOTE) are compared. The experiments use three public Kaggle datasets, with reported best accuracies of 97.5% on BT-small-2c, 99.43% on BT-large-2c, and 99.33% on BT-large-4c, and the abstract claims state-of-the-art performance.

Significance. If the evaluation protocol were clean, the paper would provide a useful large-scale benchmark: 25 feature extractors, nine classifiers, three datasets, and four ablations is a substantial empirical contribution for practitioners choosing transfer-learning pipelines for small medical image datasets. The preprocessing pipeline and the systematic hyperparameter tuning are also documented in enough detail to be reproduced. However, the headline result is not supported by the current evaluation protocol, because the same 20% test split is used both to select feature extractors, hyperparameters, and classifier ensembles and to report the final accuracy. As a result, the reported SOTA numbers are likely optimistically biased and do not estimate generalization to new patients. The paper also provides no comparison to any named prior method under identical splits, which further weakens the SOTA claim. The strength of the work is its breadth and transparency of reporting; the weakness is the absence of a valid held-out evaluation.

major comments (3)
  1. [Section 4.1.1 and Tables 4-6, 23-25] The top-2/top-3 feature extractors are selected based on the accuracies reported in Tables 4-6, but these tables report accuracy on the 20% test split defined in Section 3.2. The final ensemble results in Tables 23-25 are then reported on the same test split. This creates a test-set selection loop: the models that perform best on the test set are chosen, and their test-set accuracy is then presented as the result. The reported gains over individual models and over the claimed state of the art cannot be distinguished from selection artifact. The evaluation needs a separate validation split for model selection (or nested cross-validation), with the test set used only once for final evaluation.
  2. [Section 3.6, Eq. (15)] The hyperparameter tuning protocol is defined against the test set. Equation (15) explicitly states that HPO minimizes the loss L(X_Ts; M) on the test set, and the text says that hyperparameters producing the best outcomes on the validation set are chosen, although no validation set is introduced anywhere in Section 3.2. If hyperparameters are tuned on the same 20% test split that is later used for all reported results, then the 'fine-tuning provides a significant enhancement' claim is circular with respect to the test set. A validation split (or cross-validation) must be introduced and the HPO objective must be defined on that split.
  3. [Abstract and Section 5] The claim that the proposed method 'improve[s] upon the state of the art' is not supported by any comparison to a named prior method under identical data splits or evaluation conditions. Tables 23-25 report only internal accuracies of the proposed pipeline. To substantiate a SOTA claim, the authors need to re-implement or cite and compare with at least one or two recent baselines on the same preprocessed data and the same train/test protocol. In addition, no confidence intervals or significance tests are provided; on BT-small-2c the test set has only 51 images, so differences of one to two percentage points correspond to about one image, and across 25 feature extractors and 9 classifiers the multiple-comparison issue is substantial.
minor comments (5)
  1. [Section 3.3] The text says images are resized with bicubic interpolation but then says 'bilinear interpolation creates smoother curves'; the intended interpolation method should be stated consistently.
  2. [Section 3.4.2] The subsection heading 'vision Networks' should be 'Convolutional Neural Networks' for consistency with the rest of the paper.
  3. [Section 7] The conclusion contains a typo: 'distinguishing between brain cancer cases and and normal cases' should read 'and normal cases'.
  4. [Section 5] The paper says the top-2/top-3 ML classifiers are ensembled, but Tables 23-25 list exactly the same classifier ensembles for different feature sets; it would help to state explicitly how the ensemble members were selected and why the same three classifiers are used in all three tables.
  5. [Table 8] The 'without preprocessing' experiment uses only default hyperparameters and only CNN extractors; this is a useful sanity check, but it is not a matched ablation against the preprocessed pipeline, since the main results use tuned hyperparameters and both CNN and ViT extractors.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported SOTA accuracies are selected on the test set and then evaluated on the same test set, making the headline gains optimistically biased rather than genuine predictions.

  1. fitted input called prediction [Section 4.1.1 (Deep Feature Evaluation and Selection); Section 3.2; Tables 4-6 and 23-25]
    "we have evaluated the extracted deep features such as 12 pre-trained CNN models and 13 pre-trained ViT models using different ML classifiers as mentioned in section 3.5 and select the top three deep features for each of the three MRI datasets based on the average accuracy achieved across nine different ML classifiers"

    The paper defines only an 80/20 train/test split (Section 3.2: 'each dataset is split into a training set, comprising 80% of the total data, and a test set, making up the remaining 20%'). Section 3.6 says HPO chooses 'the best outcomes on the validation set,' but no validation set is ever described; the accuracy tables used for selection (Tables 4-6) are the test-split results. Section 4.1.1 then selects top deep features from these test-set accuracies, and the final ensemble tables (Tables 23-25) report accuracy on that same test split using the classifiers and features chosen from those very tables.

full rationale

The paper is an empirical ML study with no mathematical derivation chain, so the classical circularity modes (self-defined equations, imported uniqueness theorems) do not apply. The central claim—that the double ensemble achieves SOTA accuracy—rests on the experimental protocol. That protocol selects the top feature extractors and top classifier ensembles using test-set accuracies from Tables 4-6 (Section 4.1.1) and then reports the final ensemble numbers on the same test split in Tables 23-25. Because no independent validation split is introduced, the selection loop makes the final accuracies optimistically biased by construction: the best models on the test set are chosen and then evaluated on that same test set. This is a fitted-input-called-prediction pattern, not merely an incidental weakness. The underlying pre-trained features and classifiers are externally grounded, and there is no load-bearing self-citation, so the circularity is partial rather than total; however, the headline result as stated cannot be distinguished from selection artifact, warranting a score of 6.

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

The paper introduces no new theoretical entities, forces, or conserved quantities. Its assumptions are domain assumptions about data, transferability of pretrained features, and correct evaluation hygiene. The main free parameters are the model selection rule and classifier hyperparameters, all fit to the evaluation data in a way that is not isolated into a validation set.

free parameters (4)
  • Top-2/top-3 feature extractor selection rule = Per dataset, the top 3 models by average test accuracy
    Section 4.1.1 selects the best feature extractors using average accuracy across ML classifiers; those accuracies are computed on the test split, so the selection rule is a fitted choice that affects all ensemble results.
  • PCA component retention = Top 50% of components
    Section 5.1.2 states 'we retain top 50% of components for faster computation'; this fraction is chosen by the authors and affects the normalized-PCA ablations.
  • ML classifier hyperparameters = Grid-searched values, e.g., KNN k in 1-30, SVM C in 0.1-1000, gamma in scale/auto/0.1-10, RF n_estimators up to 500
    Table 3 defines the search spaces; the final values are selected by accuracy and used in the headline ensemble results.
  • Image augmentation schedule = Random 90-degree rotations and horizontal flips
    Section 3.3 chooses rotation and horizontal flip; the exact augmentation probability or multiplicity is not specified.
assumptions (4)
  • domain assumption ImageNet-pretrained features are informative for brain MRI classification without fine-tuning
    The entire feature extraction stage (Section 3.4.4) freezes pretrained weights and uses them as fixed feature extractors; no experiment validates the transfer from ImageNet to MRI.
  • domain assumption The 80/20 train-test split is random and data-independent
    All accuracy numbers assume the test split is a fair sample; however, feature and classifier selection is made using test-set accuracy, breaking this assumption for the reported final results.
  • domain assumption Class labels in the Kaggle datasets are correct and clinically meaningful
    The datasets are taken from Kaggle [20-22] with no manual verification or clinical annotation audit described.
  • domain assumption SMOTE is applied only to training data, not to the test set
    Section 5.1.3 describes SMOTE for balancing the datasets but does not explicitly state that the test split is excluded from oversampling; if test samples are synthetically augmented, reported accuracies are invalid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Ensemble Approaches: Optimal Deep Feature Fusion and Hyperparameter-Tuned Classifier Ensembling for Enhanced Brain Tumor Classification." pith.science (2026). https://pith.science/paper/E2ZCUI2Q

@misc{pith2026250712177,
  author       = {Pith},
  title        = {Pith review of: Hybrid Ensemble Approaches: Optimal Deep Feature Fusion and Hyperparameter-Tuned Classifier Ensembling for Enhanced Brain Tumor Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E2ZCUI2Q}},
  note         = {Machine review of arXiv:2507.12177}
}
read the original abstract

Magnetic Resonance Imaging (MRI) is widely recognized as the most reliable tool for detecting tumors due to its capability to produce detailed images that reveal their presence. However, the accuracy of diagnosis can be compromised when human specialists evaluate these images. Factors such as fatigue, limited expertise, and insufficient image detail can lead to errors. For example, small tumors might go unnoticed, or overlap with healthy brain regions could result in misidentification. To address these challenges and enhance diagnostic precision, this study proposes a novel double ensembling framework, consisting of ensembled pre-trained deep learning (DL) models for feature extraction and ensembled fine-tuned hyperparameter machine learning (ML) models to efficiently classify brain tumors. Specifically, our method includes extensive preprocessing and augmentation, transfer learning concepts by utilizing various pre-trained deep convolutional neural networks and vision transformer networks to extract deep features from brain MRI, and fine-tune hyperparameters of ML classifiers. Our experiments utilized three different publicly available Kaggle MRI brain tumor datasets to evaluate the pre-trained DL feature extractor models, ML classifiers, and the effectiveness of an ensemble of deep features along with an ensemble of ML classifiers for brain tumor classification. Our results indicate that the proposed feature fusion and classifier fusion improve upon the state of the art, with hyperparameter fine-tuning providing a significant enhancement over the ensemble method. Additionally, we present an ablation study to illustrate how each component contributes to accurate brain tumor classification.

Figures

Figures reproduced from arXiv: 2507.12177 by the authors.

Figure 1
Figure 1. Workflow of our proposed double ensembling framework for brain tumor classi [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. MRI Images with Corresponding Classes from Each Dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Pre-processing steps to crop the MRI images [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of ViT Architecture. resolution of each patch, C is the number of image channels, and N, the total number of patches, is computed as N = (HW)/P 2 . The flattened patches are then mapped to a fixed-dimensional latent vec￾tor d, referred to as patch embeddings, …
Figure 5
Figure 5. Figure 5: Our proposed double ensembling architecture for brain tumor classification. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Convolutional neural network architecture. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Methodology of transfer learning. from scratch would be impractical in this context. Instead, we employed the pre-trained, fixed weights of each ViT and CNN model, originally trained on the extensive ImageNet dataset, to effectively extract deep features from brain MRI…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Systematic Integration of Attention Modules into CNNs for Accurate and Generalizable Medical Image Diagnosis

    cs.CV 2025-09 reject novelty 3.0 of 10

    Attention-augmented CNNs usually beat plain CNNs on two medical image datasets, with EfficientNetB5 plus hybrid attention the best, but test-set-based model selection undermines the claimed consistency.

Reference graph

Works this paper leans on

63 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ullah, M

    Z. Ullah, M. Usman, M. Jeon, J. Gwak, Cascade multiscale residual at- tention cnns with adaptive roi for automatic brain tumor segmentation, Information sciences 608 (2022) 1541–1556

  2. [2]

    Popuri, D

    K. Popuri, D. Cobzas, A. Murtha, M. Jägersand, 3d variational brain tumor segmentation using dirichlet priors on a clustered feature set, Internationaljournalofcomputerassistedradiologyandsurgery7(2012) 493–506

  3. [3]

    Ullah, M

    Z. Ullah, M. U. Farooq, S.-H. Lee, D. An, A hybrid image enhancement based brain mri images classification technique, Medical hypotheses 143 (2020) 109922

  4. [4]

    Saleh, R

    A. Saleh, R. Sukaik, S. S. Abu-Naser, Brain tumor classification us- ing deep learning, in: 2020 International Conference on Assistive and Rehabilitation Technologies (iCareTech), IEEE, 2020, pp. 131–136

  5. [5]

    V. K. Waghmare, M. H. Kolekar, Brain tumor classification using deep learning, Internet of things for healthcare technologies (2021) 155–175

  6. [6]

    Aamir, Z

    M. Aamir, Z. Rahman, Z. A. Dayo, W. A. Abro, M. I. Uddin, I. Khan, A. S. Imran, Z. Ali, M. Ishfaq, Y. Guan, et al., A deep learning ap- proach for brain tumor classification using mri images, Computers and Electrical Engineering 101 (2022) 108105

  7. [7]

    Sandhiya, S

    B. Sandhiya, S. K. S. Raja, Deep learning and optimized learning ma- chine for brain tumor classification, Biomedical Signal Processing and Control 89 (2024) 105778

  8. [8]

    Agrawal, P

    T. Agrawal, P. Choudhary, V. Kumar, P. Singh, M. Diwakar, S. Kumar, A comparative study of brain tumor classification on unbalanced dataset using deep neural networks, Biomedical Signal Processing and Control 94 (2024) 106256

Show all 63 references
  1. [9]

    Priya, V

    A. Priya, V. Vasudevan, Brain tumor classification and detection via hybrid alexnet-gru based on deep learning, Biomedical Signal Processing and Control 89 (2024) 105716. 37

  2. [10]

    Deepak, P

    S. Deepak, P. Ameer, Brain tumor classification using deep cnn features via transfer learning, Computers in biology and medicine 111 (2019) 103345

  3. [11]

    A.Çinar, M.Yildirim, Detectionoftumorsonbrainmriimagesusingthe hybrid convolutional neural network architecture, Medical hypotheses 139 (2020) 109684

  4. [12]

    Khawaldeh, U

    S. Khawaldeh, U. Pervaiz, A. Rafiq, R. S. Alkhawaldeh, Noninvasive grading of glioma tumor using magnetic resonance imaging with convo- lutional neural networks, Applied Sciences 8 (2017) 27

  5. [13]

    Saxena, A

    P. Saxena, A. Maheshwari, S. Maheshwari, Predictive modeling of brain tumor: a deep learning approach, in: Innovations in Computational IntelligenceandComputerVision: ProceedingsofICICV2020, Springer, 2020, pp. 275–285

  6. [14]

    X. Yang, Y. Fan, Feature extraction using convolutional neural networks for multi-atlas based image segmentation, in: Medical Imaging 2018: Image Processing, volume 10574, SPIE, 2018, pp. 866–873

  7. [15]

    Wicht, Deep learning feature extraction for image processing, Ph.D

    B. Wicht, Deep learning feature extraction for image processing, Ph.D. thesis, éditeur non identifié, 2017

  8. [16]

    F. J. Díaz-Pernas, M. Martínez-Zarzuela, M. Antón-Rodríguez, D. González-Ortega, A deep learning approach for brain tumor clas- sification and segmentation using a multiscale convolutional neural net- work, in: Healthcare, volume 9, MDPI, 2021, p. 153

  9. [17]

    P. S. Raja, et al., Brain tumor classification using a hybrid deep au- toencoder with bayesian fuzzy clustering-based segmentation approach, Biocybernetics and Biomedical Engineering 40 (2020) 440–453

  10. [18]

    Şahin, D

    E. Şahin, D. Özdemir, H. Temurtaş, Multi-objective optimization of vit architecture for efficient brain tumor classification, Biomedical Signal Processing and Control 91 (2024) 105938

  11. [19]

    Wang, S.-Y

    J. Wang, S.-Y. Lu, S.-H. Wang, Y.-D. Zhang, Ranmerformer: Random- izedvisiontransformerwithtokenmergingforbraintumorclassification, Neurocomputing 573 (2024) 127216. 38

  12. [20]

    Bhuvaji, A

    S. Bhuvaji, A. Kadam, P. Bhumkar, S. Dedge, S. Kanchan, Brain tumor classification (mri), Kaggle 10 (2020)

  13. [21]

    Chakrabarty, Brain mri images for brain tumor detection, J

    N. Chakrabarty, Brain mri images for brain tumor detection, J. Exp. Med 216 (2019) 539–555

  14. [22]

    Hamada, Br35h brain tumor detection 2020 dataset, https://www.kaggle.com/ahmedhamada0/braintumor-detection (2020)

    A. Hamada, Br35h brain tumor detection 2020 dataset, https://www.kaggle.com/ahmedhamada0/braintumor-detection (2020)

  15. [23]

    Zhang, X

    X. Zhang, X. Zhou, M. Lin, J. Sun, Finding extreme points in contours with opencv, PyImageSearch. Available online: https://www. pyimagesearch. com/2016/04/11/finding-extreme-points- in-contours-with-opencv (accessed on 10 August 2020) (2020)

  16. [24]

    Perez, J

    L. Perez, J. Wang, The effectiveness of data augmentation in image clas- sification using deep learning, arXiv preprint arXiv:1712.04621 (2017)

  17. [25]

    S. Yang, W. Xiao, M. Zhang, S. Guo, J. Zhao, F. Shen, Image data aug- mentation for deep learning: A survey, arXiv preprint arXiv:2204.08610 (2022)

  18. [26]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

  19. [27]

    Wu, Introduction to convolutional neural networks, National Key Lab for Novel Software Technology

    J. Wu, Introduction to convolutional neural networks, National Key Lab for Novel Software Technology. Nanjing University. China 5 (2017) 495

  20. [28]

    Goyal, R

    M. Goyal, R. Goyal, B. Lall, Learning activation functions: A new paradigm for understanding neural networks, arXiv preprint arXiv:1906.09529 (2019)

  21. [29]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in neural information processing systems 25 (2012)

  22. [30]

    Akçay, M

    S. Akçay, M. E. Kundegorski, M. Devereux, T. P. Breckon, Trans- fer learning using convolutional neural networks for object classification within x-ray baggage security imagery, in: 2016 IEEE international conference on image processing (ICIP), IEEE, 2016, pp. 1057–1061. 39

  23. [31]

    I. M. Baltruschat, H. Nickisch, M. Grass, T. Knopp, A. Saalbach, Com- parison of deep learning approaches for multi-label chest x-ray classifi- cation, Scientific reports 9 (2019) 6381

  24. [32]

    Christodoulidis, M

    S. Christodoulidis, M. Anthimopoulos, L. Ebner, A. Christe, S. Mougiakakou, Multisource transfer learning with convolutional neu- ral networks for lung pattern analysis, IEEE journal of biomedical and health informatics 21 (2016) 76–84

  25. [33]

    Tajbakhsh, J

    N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B. Gotway, J. Liang, Convolutional neural networks for medical image analysis: Full training or fine tuning?, IEEE transactions on medical imaging 35 (2016) 1299–1312

  26. [34]

    Weiss, T

    K. Weiss, T. M. Khoshgoftaar, D. Wang, A survey of transfer learning, Journal of Big data 3 (2016) 1–40

  27. [35]

    B. Wu, C. Xu, X. Dai, A. Wan, P. Zhang, Z. Yan, M. Tomizuka, J. Gon- zalez, K. Keutzer, P. Vajda, Visual transformers: Token-based im- age representation and processing for computer vision, arXiv preprint arXiv:2006.03677 (2020)

  28. [36]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, 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

  29. [37]

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

  30. [38]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely con- nected convolutional networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4700–4708

  31. [39]

    Simonyan, A

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

  32. [40]

    Krizhevsky, One weird trick for parallelizing convolutional neural networks, arXiv preprint arXiv:1404.5997 (2014)

    A. Krizhevsky, One weird trick for parallelizing convolutional neural networks, arXiv preprint arXiv:1404.5997 (2014). 40

  33. [41]

    Szegedy, V

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2818–2826

  34. [42]

    S. Xie, R. Girshick, P. Dollár, Z. Tu, K. He, Aggregated residual trans- formationsfordeepneuralnetworks, in: ProceedingsoftheIEEEconfer- ence on computer vision and pattern recognition, 2017, pp. 1492–1500

  35. [43]

    Sandler, A

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.-C. Chen, Mo- bilenetv2: Inverted residuals and linear bottlenecks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520

  36. [44]

    M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, Q. V. Le, Mnasnet: Platform-aware neural architecture search for mobile, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2820–2828

  37. [45]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Scikit-learn: Machine learning in python, the Journal of machine Learn- ing research 12 (2011) 2825–2830

  38. [46]

    X. Meng, J. Jiang, H. Wang, Agwo: Advanced gwo in multi-layer per- ception optimization, Expert Systems with Applications 173 (2021) 114676

  39. [47]

    Freund, R

    Y. Freund, R. E. Schapire, A decision-theoretic generalization of on- line learning and an application to boosting, Journal of computer and system sciences 55 (1997) 119–139

  40. [48]

    Breiman, Random forests, Machine learning 45 (2001) 5–32

    L. Breiman, Random forests, Machine learning 45 (2001) 5–32

  41. [49]

    Cortes, V

    C. Cortes, V. Vapnik, Support-vector networks, Machine learning 20 (1995) 273–297

  42. [50]

    T. Yu, H. Zhu, Hyper-parameter optimization: A review of algorithms and applications, arXiv preprint arXiv:2003.05689 (2020). 41

  43. [51]

    Tran, J.-G

    N. Tran, J.-G. Schneider, I. Weber, A. K. Qin, Hyper-parameter opti- mization in classification: To-do or not-to-do, Pattern Recognition 103 (2020) 107245

  44. [52]

    Claesen, B

    M. Claesen, B. De Moor, Hyperparameter search in machine learning, arXiv preprint arXiv:1502.02127 (2015)

  45. [53]

    D. M. Belete, M. D. Huchaiah, Grid search in hyperparameter optimiza- tion of machine learning models for prediction of hiv/aids test results, International Journal of Computers and Applications 44 (2022) 875–886

  46. [54]

    Rajagopal, P

    S. Rajagopal, P. P. Kundapur, K. S. Hareesha, A stacking ensemble for network intrusion detection using heterogeneous datasets, Security and Communication Networks 2020 (2020) 4586875

  47. [55]

    Henderi, T

    H. Henderi, T. Wahyuningsih, E. Rahwanto, Comparison of min-max normalization and z-score normalization in the k-nearest neighbor (knn) algorithm to test the accuracy of types of breast cancer, International Journal of Informatics and Information Systems 4 (2021) 13–20

  48. [56]

    Singh, B

    D. Singh, B. Singh, Feature wise normalization: An effective way of normalizing data, Pattern Recognition 122 (2022) 108307

  49. [57]

    J. Han, J. Pei, H. Tong, Data mining: concepts and techniques, Morgan kaufmann, 2022

  50. [58]

    N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, Smote: synthetic minority over-sampling technique, Journal of artificial intelli- gence research 16 (2002) 321–357

  51. [59]

    D. A. Cieslak, N. V. Chawla, A. Striegel, Combating imbalance in network intrusion datasets., in: GrC, 2006, pp. 732–737

  52. [60]

    Fallahi, S

    A. Fallahi, S. Jafari, An expert system for detection of breast cancer using data preprocessing and bayesian network, International Journal of Advanced Science and Technology 34 (2011) 65–70

  53. [61]

    A. Saha, A. N. Chaudhury, P. Bhowmik, R. Chatterjee, Awareness of cervical cancer among female students of premier colleges in kolkata, india, Asian Pac J Cancer Prev 11 (2010) 1085–90. 42

  54. [62]

    Vida, Practical guide to deal with imbalanced classification prob- lems in r

    A. Vida, Practical guide to deal with imbalanced classification prob- lems in r. imbalanceanalyticsvida, URl: https://www. analyticsvid- hya. com/blog/2016/03/practical-guide-deal-imbalanced-classification- problems/(visited on 08/07/2017) (2016)

  55. [63]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, A. Courville, Y. Bengio, Deep learning, vol- ume 1, MIT press Cambridge, 2016. 43

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.