pith. sign in

arxiv: 2509.06713 · v1 · submitted 2025-09-08 · 💻 cs.CV · cs.AI

MRI-Based Brain Tumor Detection through an Explainable EfficientNetV2 and MLP-Mixer-Attention Architecture

Mustafa Yurdakul , \c{S}akir Ta\c{s}demir This is my paper

Pith reviewed 2026-05-18 18:15 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords brain tumor classificationMRIEfficientNetV2MLP-Mixerattention mechanismGrad-CAMdeep learningexplainable AI
0
0 comments X

The pith

Combining EfficientNetV2 with an attention-based MLP-Mixer classifies brain tumors from MRI at 99.50 percent accuracy.

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

The paper first benchmarks nine standard CNN architectures on the public Figshare collection of 3064 T1-weighted contrast-enhanced MRI images and identifies EfficientNetV2 as the strongest single backbone. An attention-based MLP-Mixer module is then attached to this backbone to improve feature mixing across spatial patches before final classification into three tumor categories. The complete model is assessed with five-fold cross-validation, yielding 99.50 percent accuracy, 99.47 percent precision, 99.52 percent recall and 99.49 percent F1 score while also generating Grad-CAM heatmaps. These visualizations are presented as evidence that the network attends to tumor locations rather than extraneous image regions. The authors position the resulting system as both more accurate than prior published approaches on the same data and sufficiently interpretable for clinical decision support.

Core claim

By selecting EfficientNetV2 after evaluating nine CNNs on the Figshare dataset and integrating an attention-based MLP-Mixer, the authors obtain a hybrid architecture that reaches 99.50 percent accuracy, 99.47 percent precision, 99.52 percent recall and 99.49 percent F1 score under five-fold cross-validation, exceeds previously reported results on the identical collection of 3064 images, and produces Grad-CAM maps that align with clinically relevant tumor regions.

What carries the argument

Attention-based MLP-Mixer module attached to an EfficientNetV2 backbone to refine spatial feature mixing and support visual explanation via Grad-CAM.

If this is right

  • The hybrid model surpasses both plain CNN baselines and previously published methods on the Figshare dataset.
  • Grad-CAM visualizations confirm that decisions rest on tumor locations rather than artifacts.
  • Five-fold cross-validation produces consistent metrics above 99.4 percent across all folds.
  • The architecture supplies both high numerical accuracy and visual interpretability required for clinical decision support.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the backbone was selected after inspecting results on the evaluation data, the quoted accuracy may be optimistically biased relative to truly unseen clinical scans.
  • The same EfficientNetV2-plus-attention-MLP-Mixer pattern could be tested on other MRI tasks such as segmentation or on non-brain tumor types, though no such experiments appear in the paper.
  • Robustness to variations in field strength, contrast protocols, or patient demographics outside the Figshare collection remains untested.

Load-bearing premise

Choosing the single highest-performing CNN backbone from nine candidates evaluated on the same dataset before adding the MLP-Mixer module will yield performance estimates that generalize to new clinical MRI scans.

What would settle it

Testing the trained model on an independent set of MRI scans acquired from different hospitals or scanner vendors would show whether accuracy falls below 95 percent.

Figures

Figures reproduced from arXiv: 2509.06713 by \c{S}akir Ta\c{s}demir, Mustafa Yurdakul.

Figure 1
Figure 1. Figure 1: Schematic diagram of proposed study 2.1. Dataset The Figshare dataset[24] used in this study was created from MRI images obtained during brain tumor examinations performed between 2005 and 2010 at Nanfang Hospital and Tianjin Medical University General Hospital in China. It includes 3064 T1-weighted contrast-enhanced brain MRI images from a total of 233 patients. The dataset was first published online in 2… view at source ↗
Figure 2
Figure 2. Figure 2: MRI slices from the brain tumor dataset for glioma (top), meningioma (middle), and pituitary tumor (bottom). Each row shows axial (first two), coronal (middle two), and sagittal (last two) views. 2.2. Linear Attention Linear attention[25] is a variant of the attention mechanism designed to efficiently capture context dependencies in long sequences with reduced computational cost. The basic principle is to … view at source ↗
Figure 3
Figure 3. Figure 3: Schematic representation of the proposed MLP-Mixer Attention block 2.4. Proposed methodology EfficientNetV2[27] was selected as the backbone of the model because it provides an optimal balance between high accuracy and computational efficiency. EfficientNetV2 has different scales (depth, width, and resolution). A small scale was selected for this study [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic representation of the proposed EfficientNetV2–MLP Mixer Attention model 2. 5. Explainability with Grad-CAM Grad-CAM[28] is a method used to visualize how DL models make predictions. It aims to find out which areas of the image the model pays more attention to when making decisions. First, the score function 𝑦 𝑐 of the target class is calculated, and the derivatives of this score on each feature m… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of DL models across performance metrics In conclusion, the findings reveal that CNN-based architectures provide a robust framework for brain tumor classification. The comparison between the models shows that EfficientNetV2 stands out in particular for achieving balanced results in terms of both high accuracy and other metrics such as precision, recall, and F1-score. Therefore, the preference for… view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices of the EfficientNetV2 and proposed model(0: meningioma, 1: glioma, 2: pituitary tumor) The confusion matrices presented in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative MRI slices of brain tumors: (a) original T1-weighted contrast-enhanced images, (b) tumor regions annotated with ground truth, and (c) Grad-CAM heatmaps highlighting the model’s focus areas for classification in glioma, meningioma, and pituitary tumors. When [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
read the original abstract

Brain tumors are serious health problems that require early diagnosis due to their high mortality rates. Diagnosing tumors by examining Magnetic Resonance Imaging (MRI) images is a process that requires expertise and is prone to error. Therefore, the need for automated diagnosis systems is increasing day by day. In this context, a robust and explainable Deep Learning (DL) model for the classification of brain tumors is proposed. In this study, a publicly available Figshare dataset containing 3,064 T1-weighted contrast-enhanced brain MRI images of three tumor types was used. First, the classification performance of nine well-known CNN architectures was evaluated to determine the most effective backbone. Among these, EfficientNetV2 demonstrated the best performance and was selected as the backbone for further development. Subsequently, an attention-based MLP-Mixer architecture was integrated into EfficientNetV2 to enhance its classification capability. The performance of the final model was comprehensively compared with basic CNNs and the methods in the literature. Additionally, Grad-CAM visualization was used to interpret and validate the decision-making process of the proposed model. The proposed model's performance was evaluated using the five-fold cross-validation method. The proposed model demonstrated superior performance with 99.50% accuracy, 99.47% precision, 99.52% recall and 99.49% F1 score. The results obtained show that the model outperforms the studies in the literature. Moreover, Grad-CAM visualizations demonstrate that the model effectively focuses on relevant regions of MRI images, thus improving interpretability and clinical reliability. A robust deep learning model for clinical decision support systems has been obtained by combining EfficientNetV2 and attention-based MLP-Mixer, providing high accuracy and interpretability in brain tumor classification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a hybrid deep learning model for brain tumor classification in T1-weighted contrast-enhanced MRI images from the public Figshare dataset (3064 images across three tumor classes). It first benchmarks nine CNN backbones, selects EfficientNetV2 as the best performer, integrates an attention-based MLP-Mixer module, and reports 99.50% accuracy, 99.47% precision, 99.52% recall, and 99.49% F1-score under 5-fold cross-validation. The model is positioned as outperforming prior literature methods while providing clinical interpretability through Grad-CAM visualizations.

Significance. If the performance claims can be substantiated without selection bias and with proper external validation, the work would offer a useful architectural combination of EfficientNetV2 and attention-enhanced MLP-Mixer for medical image classification, together with a positive emphasis on explainability. The Grad-CAM analysis is a constructive element for clinical relevance. However, the current evaluation setup on a single modest-sized public dataset limits the strength of the superiority claim for real-world deployment.

major comments (2)
  1. [Abstract / Methods] Abstract and Methods (backbone selection procedure): Evaluating all nine CNN architectures on the full Figshare dataset to select EfficientNetV2, then attaching the MLP-Mixer module and reporting 5-fold CV metrics on the same data distribution, performs model selection on the evaluation set. This risks optimistic bias in the headline 99.50% accuracy / 99.49% F1 figures; no nested CV, separate validation split for ranking, or standalone performance of the selected backbone is described.
  2. [Results] Results / Experimental Setup: No external test set, multi-center cohort, or independent clinical validation data is used. All metrics derive from 5-fold CV on the single 3064-image Figshare collection; this is a load-bearing gap for any claim of clinical reliability or outperformance of literature methods on unseen MRI data.
minor comments (2)
  1. [Abstract] The abstract and methods should explicitly state whether backbone selection occurred inside or outside the 5-fold CV loops and provide the per-backbone metrics that justified choosing EfficientNetV2.
  2. [Methods] Training details (optimizer, learning rate schedule, data augmentation, batch size, and early-stopping criteria) are insufficiently specified for reproducibility of the reported 99.5% metrics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the detailed and constructive feedback on our manuscript. We have carefully reviewed the comments and provide point-by-point responses below. Revisions have been made to the manuscript to enhance methodological transparency and to moderate claims about generalizability, thereby addressing the raised concerns.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and Methods (backbone selection procedure): Evaluating all nine CNN architectures on the full Figshare dataset to select EfficientNetV2, then attaching the MLP-Mixer module and reporting 5-fold CV metrics on the same data distribution, performs model selection on the evaluation set. This risks optimistic bias in the headline 99.50% accuracy / 99.49% F1 figures; no nested CV, separate validation split for ranking, or standalone performance of the selected backbone is described.

    Authors: We thank the referee for highlighting this important methodological consideration. The backbone selection was conducted using the same 5-fold cross-validation protocol applied to all nine architectures to ensure fair and consistent comparison under identical conditions. To improve transparency, we have revised the Methods section to detail this procedure and added Table 2 reporting the full performance metrics (accuracy, precision, recall, F1) for every backbone under the identical 5-fold CV. This allows independent verification of the selection rationale. While a nested CV would offer stricter separation, our protocol follows common practice in comparative architecture studies on this benchmark; we have added an explicit limitations paragraph discussing the risk of optimistic bias. revision: partial

  2. Referee: [Results] Results / Experimental Setup: No external test set, multi-center cohort, or independent clinical validation data is used. All metrics derive from 5-fold CV on the single 3064-image Figshare collection; this is a load-bearing gap for any claim of clinical reliability or outperformance of literature methods on unseen MRI data.

    Authors: We agree that external validation on independent multi-center data would strengthen claims of clinical reliability. The Figshare dataset remains the standard public benchmark used by the majority of prior brain-tumor MRI classification studies, permitting direct apples-to-apples comparison with the literature. Our 5-fold CV results are therefore reported in the same evaluation regime as those works. In the revised manuscript we have (i) tempered language regarding real-world deployment, (ii) explicitly framed the results as benchmark performance on this single public collection, and (iii) added a forward-looking statement calling for future multi-center validation. Grad-CAM visualizations provide additional qualitative evidence that the model attends to clinically relevant regions. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results on public dataset with standard model selection

full rationale

The paper reports measured classification accuracies (99.50% etc.) obtained via 5-fold cross-validation on the fixed public Figshare dataset after selecting EfficientNetV2 as the best of nine CNN backbones evaluated on that same data and then attaching an MLP-Mixer attention module. No equation, parameter, or claimed prediction reduces by construction to a fitted input or self-citation; the central claims are direct empirical measurements against external benchmarks rather than a derivation that is definitionally equivalent to its inputs. The backbone-selection step introduces a methodological risk of optimistic bias but does not constitute any of the enumerated circularity patterns (self-definitional, fitted-input-called-prediction, self-citation load-bearing, etc.). The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claim rests on empirical fitting of a large neural network whose weights and many training choices are free parameters optimized on the target dataset; the only explicit domain assumption is that the Figshare collection is a suitable benchmark.

free parameters (1)
  • Backbone selection and all model hyperparameters
    EfficientNetV2 chosen after evaluating nine architectures on the same data; numerous learning-rate, batch-size, and architecture-specific parameters fitted to maximize reported metrics.
axioms (1)
  • domain assumption The Figshare dataset of 3064 T1-weighted contrast-enhanced images is representative for evaluating brain-tumor classification performance.
    All quantitative claims and comparisons to literature depend on this dataset being a valid proxy for clinical generalization.

pith-pipeline@v0.9.0 · 5858 in / 1507 out tokens · 46547 ms · 2026-05-18T18:15:13.669199+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 2 internal anchors

  1. [1]

    Shanthakumar, P. and P. Ganeshkumar, Performance analysis of classifier for brain tumor detection and diagnosis. Computers & Electrical Engineering, 2015. 45: p. 302-311

    work page 2015

  2. [2]

    Rajan, P. and C. Sundar, Brain tumor detection and segmentation by intensity adjustment. Journal of medical systems, 2019. 43(8): p. 282

    work page 2019

  3. [3]

    JAMA network open, 2021

    Rahib, L., et al., Estimated projection of US cancer incidence and death to 2040. JAMA network open, 2021. 4(4): p. e214708-e214708

    work page 2040

  4. [4]

    McFaline-Figueroa, J.R. and E.Q. Lee, Brain tumors. The American journal of medicine, 2018. 131(8): p. 874-882

    work page 2018

  5. [5]

    Nature Reviews Disease Primers, 2024

    Weller, M., et al., Glioma. Nature Reviews Disease Primers, 2024. 10(1): p. 33

    work page 2024

  6. [6]

    Nature clinical practice Neurology, 2006

    Schwartzbaum, J.A., et al., Epidemiology and molecular pathology of glioma. Nature clinical practice Neurology, 2006. 2(9): p. 494-503

    work page 2006

  7. [7]

    Volume 2: The Path to Bedside Management

    Bailo, M., et al., Meningioma and other meningeal tumors, in Human Brain and Spinal Cord Tumors: From Bench to Bedside. Volume 2: The Path to Bedside Management. 2023, Springer. p. 73-97

    work page 2023

  8. [8]

    Fathi, A.-R. and U. Roelcke, Meningioma. Current neurology and neuroscience reports, 2013. 13(4): p. 337

    work page 2013

  9. [9]

    Dworakowska, D. and A.B. Grossman, Aggressive and malignant pituitary tumours: state-of- the-art. Endocrine-related cancer, 2018. 25(11): p. R559-R575

    work page 2018

  10. [10]

    Berrocal, and E

    Araujo-Castro, M., V.R. Berrocal, and E. Pascual-Corrales, Pituitary tumors: epidemiology and clinical presentation spectrum. Hormones, 2020. 19(2): p. 145-155

    work page 2020

  11. [11]

    Journal of Imaging Informatics in Medicine, 2025: p

    Jin, S., et al., Preoperative Prediction of Non-functional Pituitary Neuroendocrine Tumors and Posterior Pituitary Tumors Based on MRI Radiomic Features. Journal of Imaging Informatics in Medicine, 2025: p. 1-12

    work page 2025

  12. [12]

    Archives of computational methods in engineering, 2022

    Ali, S., et al., A comprehensive survey on brain tumor diagnosis using deep learning and emerging hybrid techniques with multi-modal MR image. Archives of computational methods in engineering, 2022. 29(7): p. 4871-4896

    work page 2022

  13. [13]

    Expert systems with Applications, 2014

    El-Dahshan, E.-S.A., et al., Computer-aided diagnosis of human brain tumor through MRI: A survey and a new algorithm. Expert systems with Applications, 2014. 41(11): p. 5526-5545

    work page 2014

  14. [14]

    Olabarriaga, S.D. and A.W. Smeulders, Interaction in the segmentation of medical images: A survey. Medical image analysis, 2001. 5(2): p. 127-142

    work page 2001

  15. [15]

    Computerized Medical Imaging and Graphics, 2019

    Swati, Z.N.K., et al., Brain tumor classification for MR images using transfer learning and fine-tuning. Computerized Medical Imaging and Graphics, 2019. 75: p. 34-46

    work page 2019

  16. [16]

    Computers and Electrical Engineering, 2022

    Aamir, M., et al., A deep learning approach for brain tumor classification using MRI images. Computers and Electrical Engineering, 2022. 101: p. 108105

    work page 2022

  17. [17]

    Mohammed, and H

    Ismael, S.A.A., A. Mohammed, and H. Hefny, An enhanced deep learning approach for brain cancer MRI images classification using residual networks. Artificial intelligence in medicine,

    work page

  18. [18]

    biocybernetics and biomedical engineering, 2020

    Hashemzehi, R., et al., Detection of brain tumors from MRI images base on deep learning using hybrid model CNN and NADE. biocybernetics and biomedical engineering, 2020. 40(3): p. 1225-1232

    work page 2020

  19. [19]

    Biomedical Signal Processing and Control, 2023

    Sharma, A.K., et al., Brain tumor classification using the modified ResNet50 model based on transfer learning. Biomedical Signal Processing and Control, 2023. 86: p. 105299

    work page 2023

  20. [20]

    Khaliki, M.Z. and M.S. Başarslan, Brain tumor detection from images and comparison with transfer learning methods and 3-layer CNN. Scientific Reports, 2024. 14(1): p. 2664

    work page 2024

  21. [21]

    Biomedical Signal Processing and Control,

    Khan, S.U.R., et al., ShallowMRI: A novel lightweight CNN with novel attention mechanism for Multi brain tumor classification in MRI images. Biomedical Signal Processing and Control,

    work page

  22. [22]

    Computers in Biology and Medicine,

    Shaikh, A., et al., Enhanced brain tumor detection and segmentation using densely connected convolutional networks with stacking ensemble learning. Computers in Biology and Medicine,

    work page

  23. [23]

    Davar, S. and T. Fevens, Enhanced U -Net Architecture for Brain Tumour Localization & Segmentation in T1 -Weighted MRI. IEEE Transactions on Circuits and Systems II: Express Briefs, 2025

    work page 2025

  24. [24]

    (No Title), 2015

    Cheng, J., Brain tumor dataset. (No Title), 2015

    work page 2015

  25. [25]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Katharopoulos, A., et al. Transformers are rnns: Fast autoregressive transformers with linear attention. in International conference on machine learning. 2020. PMLR

    work page 2020

  26. [26]

    Advances in neural information processing systems, 2021

    Tolstikhin, I.O., et al., Mlp-mixer: An all -mlp architecture for vision. Advances in neural information processing systems, 2021. 34: p. 24261-24272

    work page 2021

  27. [27]

    Tan, M. and Q. Le. Efficientnetv2: Smaller models and faster training . in International conference on machine learning. 2021. PMLR

    work page 2021

  28. [28]

    Grad-cam: Visual explanations from deep networks via gradient -based localization

    Selvaraju, R.R., et al. Grad-cam: Visual explanations from deep networks via gradient -based localization. in Proceedings of the IEEE international conference on computer vision. 2017

    work page 2017

  29. [29]

    A convnet for the 2020s

    Liu, Z., et al. A convnet for the 2020s. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022

    work page 2022

  30. [30]

    Densely connected convolutional networks

    Huang, G., et al. Densely connected convolutional networks . in Proceedings of the IEEE conference on computer vision and pattern recognition. 2017

    work page 2017

  31. [31]

    Inceptionnext: When inception meets convnext

    Yu, W., et al. Inceptionnext: When inception meets convnext . in Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. 2024

    work page 2024

  32. [32]

    MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

    Howard, A.G., et al., Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    Resnest: Split-attention networks

    Zhang, H., et al. Resnest: Split-attention networks. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022

    work page 2022

  34. [34]

    Deep residual learning for image recognition

    He, K., et al. Deep residual learning for image recognition . in Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    work page 2016

  35. [35]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  36. [36]

    Xception: Deep learning with depthwise separable convolutions

    Chollet, F. Xception: Deep learning with depthwise separable convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2017

    work page 2017