CoChat
Sign Up
Cited by 124Digital Imaging for Blood DiseasesAI in cancer detectionCell Image Analysis Techniques

Method for diagnosis of acute lymphoblastic leukemia based on vit-cnn ensemble model

Open access
Jiang, Zhencun, Dong, Zhengxin, Wang, Lingyang, Jiang, Wenping·2021·Computational Intelligence and Neuroscience
DOI 10.1155/2021/7529893PMC PMC8405335
Open this paper in CoChat — free accountOpen PDF

Abstract

Acute lymphocytic leukemia (ALL) is a deadly cancer that not only affects adults but also accounts for about 25% of childhood cancers. Timely and accurate diagnosis of the cancer is an important premise for effective treatment to improve survival rate. Since the image of leukemic B‐lymphoblast cells (cancer cells) under the microscope is very similar in morphology to that of normal B‐lymphoid precursors (normal cells), it is difficult to distinguish between cancer cells and normal cells. Therefore, we propose the ViT‐CNN ensemble model to classify cancer cells images and normal cells images to assist in the diagnosis of acute lymphoblastic leukemia. The ViT‐CNN ensemble model is an ensemble model that combines the vision transformer model and convolutional neural network (CNN) model. The vision transformer model is an image classification model based entirely on the transformer structure, which has completely different feature extraction method from the CNN model. The ViT‐CNN ensemble model can extract the features of cells images in two completely different ways to achieve better classification results. In addition, the data set used in this article is an unbalanced data set and has a certain amount of noise, and we propose a difference enhancement‐random sampling (DERS) data enhancement method, create a new balanced data set, and use the symmetric cross‐entropy loss function to reduce the impact of noise in the data set. The classification accuracy of the ViT‐CNN ensemble model on the test set has reached 99.03%, and it is proved through experimental comparison that the effect is better than other models. The proposed method can accurately distinguish between cancer cells and normal cells and can be used as an effective method for computer‐aided diagnosis of acute lymphoblastic leukemia.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a ViT-CNN ensemble model combining a vision transformer and EfficientNet-b0 to classify leukemic B-lymphoblast cells versus normal B-lymphoid precursors for acute lymphoblastic leukemia diagnosis. On the ISBI 2019 dataset, the ensemble reached 99.03% accuracy and 99.14% precision.

Key findings

  • The ViT-CNN ensemble achieved 99.03% accuracy and 99.14% precision on the test set, outperforming the vision transformer and EfficientNet models individually.
  • The vision transformer model’s accuracy was 3.72% higher and precision 1.52% higher than the EfficientNet model.
  • The ensemble’s accuracy was 4.08%, 5.38%, and 3.79% higher than ResNet50, DenseNet121, and VGG16, respectively.
  • Compared with the model in literature [11], which had 96.58% accuracy, the proposed ensemble’s 99.03% accuracy was 2.45% higher.

Methods

  • Used the ISBI 2019 dataset with 10,661 cell images from 73 subjects: 7,272 cancer cell images from 47 ALL patients and 3,389 normal cell images from 26 healthy persons.
  • Proposed difference enhancement-random sampling (DERS) to address class imbalance, generating 29,088 cancer cell images and 23,702 normal cell images, then randomly drawing 20,000 images to create a balanced dataset; images were resized to 224 × 224 and normalized.
  • Trained a vision transformer and EfficientNet-b0 via transfer learning fine-tuning for 30 epochs with learning rate 0.0001 and batch size 32, using symmetric cross-entropy loss and RAdam optimizer; combined outputs by weighted sum with ViT coefficient 0.7 and EfficientNet coefficient 0.3, then evaluated accuracy and precision on an 8:1:1 train/validation/test split.

Limitations

  • The paper does not state a limitations section.
  • Evident from the text: evaluation uses only one dataset (ISBI 2019) and no external validation is reported.
  • Evident from the text: performance is assessed only with accuracy and precision, with no sensitivity, specificity, or recall reported.
  • Evident from the text: the DERS method relies on geometric transformations and random sampling, which may not capture real clinical image variation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1155/2021/7529893

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The vision transformer model has a 3.72% higher accuracy rate than the EfficientNet model on the test set.

    “As can be seen from Table 1, the accuracy rate of the vision transformer model is 3.72% higher than that of the EfficientNet model”
    Quote in the paperSubstantial claim
  • Findingown data

    The ViT-CNN ensemble model has a 3.79% higher accuracy rate than VGG16.

    “It can be seen from Table 2 that the accuracy rate of the ViT-CNN ensemble model is 4.08%, 5.38%, and 3.79% higher than the accuracy rate of the Resnet50, Densenet121, and VGG16.”
    Quote in the paperSubstantial claim
  • Findingnone given

    The morphology of leukemic B-lymphoblast cells (cancer cells) and normal B-lymphoid precursors (normal cells) under the microscope is very similar, making them difficult to distinguish.

    “Since the image of leukemic B-lymphoblast cells (cancer cells) under the microscope is very similar in morphology to that of normal B-lymphoid precursors (normal cells), it is difficult to distinguish between cancer cells and normal cells.”
    • Also stated in: § 2.1. Data Set · plain
      “As shown in Figure 1, the morphology of the two cells is very similar, so a professional oncologist will annotate the label of the image.”
    Quote in the paperSubstantial claim
  • Findingown data

    The vision transformer model has a 1.52% higher precision rate than the EfficientNet model on the test set.

    “and the precision rate of the vision transformer model is 1.52% higher than that of the EfficientNet model”
    Quote in the paperSubstantial claim
  • Findingown data

    The ViT-CNN ensemble model has a stronger ability to recognize cancer cells than the vision transformer model.

    “The ViT-CNN ensemble model has the same recognition accuracy for normal cells as the vision transformer model but has a stronger ability to recognize cancer cells.”
    Quote in the paperSubstantial claim
  • May combine two claims. 11
  • The quote only partly states this claim. 2
  • Classified as probably not a substantial claim. 7

  • May combine two claims.

    Findingcitation

    The ISBI 2019 dataset used contains 10,661 cell images from 73 subjects, including 7,272 leukemic B-lymphoblast (cancer) cell images from 47 patients and 3,389 normal B-lymphoid precursor cell images from 26 healthy persons.

    “The data set used to build the diagnostic model is the ISBI 2019 data set [16–20], and 10661 cell pictures of 73 subjects were selected, including 7272 pictures of leukemic B-lymphoblast cells (cancer cells) from 47 all patients and 3389 pictures of B-lymphoid precursors (normal cells) from 26 healthy persons.”
    Quote in the paperSubstantial claim0 of 5 sources checked
    • cites [16] Gupta R. Mallick P. Duggal R. Gupta A. Sharma O. Stain color normalization and segmentation of plasma cells in microscopic images as a prelude to development of computer assisted automated disease diagnostic tool in multiple myeloma Proceedings of the 16th International Myeloma Workshop (IMW) March 2017 New Delhi, India 10.1016/j.clml.2017.03.178 source

      Not checked in the free run.

    • cites [17] Duggal R. Gupta A. Gupta R. Wadhwa M. Ahuja C. Overlapping cell nuclei segmentation in microscopic images using deep belief networks Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP) December 2016 Guwahati, India source

      Not checked in the free run.

    • cites [18] Duggal R. Gupta A. Gupta R. Segmentation of overlapping/touching white blood cell nuclei using artificial neural networks Proceedings of the CME Series on Hemato-Oncopathology July 2016 New Delhi, India All India Institute of Medical Sciences (AIIMS)

      Not checked in the free run.

    • cites [19] Duggal R. Gupta A. Gupta R. Mallick P. Descoteaux M. Maier-Hein L. Franz A. Jannin P. Collins D. Duchesne S. SD-layer: stain deconvolutional layer for CNNs in medical microscopic imaging Proceedings of the Medical Image Computing and Computer-Assisted Intervention−MICCAI 2017 September 2017 Quebec City, Canada Springer 435 443 10.1007/978-3-319-66179-7_50 2-s2.0-85029513084 source

      Not checked in the free run.

    • cites [20] Buslaev A. Iglovikov V. I. Khvedchenya E. Parinov A. Druzhinin M. Kalinin A. A. Albumentations: fast and flexible image augmentations Information 2020 11 2 p. 125 10.3390/info11020125 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The new dataset created by DERS contains 20,000 images randomly drawn from the newly generated images and is completely balanced.

    “A new data set is created with 20,000 images randomly drawn from two newly generated images. The new data set is a completely balanced data set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The EfficientNet model has a difference in the ability to recognize cancer cells and normal cells, and the ability to recognize cancer cells is stronger.

    “The EfficientNet model has a difference in the ability to recognize cancer cells and normal cells, and the ability to recognize cancer cells is stronger.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    The DERS data enhancement method creates a new balanced dataset from the unbalanced ISBI 2019 dataset.

    “We use the method of difference enhancement-random sampling to process the data set.”
    • Also stated in: § 2.3. Data Processing · plain
      “A new data set is created with 20,000 images randomly drawn from two newly generated images. The new data set is a completely balanced data set.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Directly training on an unbalanced data set can easily cause the model to fall into overfitting or cause the generalization ability of the model to be weak.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Accurately distinguishing B-lymphoid precursors (normal cells) from cancer cells is the key to the diagnosis of acute lymphoblastic leukemia.

    “Accurately distinguishing B-lymphoid precursors (normal cells) from cancer cells is the key to the diagnosis of acute lymphoblastic leukemia.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The main diagnostic method for acute lymphoblastic leukemia is through morphology, and when there are a large number of B-lymphoblast cells in the bone marrow, it can be diagnosed as acute lymphoblastic leukemia.

    “In general, the main diagnostic method for acute lymphoblastic leukemia is through morphology. When there are a large number of B-lymphoblast cells (cancer cells) in the bone marrow, it can be diagnosed as acute lymphoblastic leukemia [3].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Peng X. Yang W. Jia L. Application of cell morphology-related technology in hematological tumors Chinese Journal of Biological Engineering 2019 39 9 84 90

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    In practical applications, better identification of cancer cells can more accurately diagnose acute lymphocytic leukemia, so the ViT-CNN ensemble model can in a superior way assist in the diagnosis of acute lymphocytic leukemia.

    “In practical applications, better identification of cancer cells can more accurately diagnose acute lymphocytic leukemia, so the ViT-CNN ensemble model can in a superior way assist in the diagnosis of acute lymphocytic leukemia.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    The ensemble model is a method to improve the accuracy rate of the model.

    “The ensemble model is a method to improve the accuracy rate of the model [26, 27].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [26] Ahmad F. Farooq A. Ghani M. U. Deep ensemble model for classification of novel coronavirus in chest X-ray images Computational Intelligence and Neuroscience 2021 2021 17 8890226 10.1155/2021/8890226 PMC7805527 33488691 source

      Not checked in the free run.

    • cites [27] Jin L.-peng Dong J. Ensemble deep learning for biomedical time series classification Computational Intelligence and Neuroscience 2016 2016 13 6212684 10.1155/2016/6212684 2-s2.0-84990841514 PMC5048093 27725828 source

      Not checked in the free run.

    • Rests on an assumption: The greater the differences between the models, the greater the performance improvement after the ensemble.

  • May combine two claims.

    Premisecitation

    A pretraining model obtained by training on large data sets can be trained with a few data sets, and the training time required for deep learning is greatly shortened.

    “The pretraining model obtained by training on large data sets can be trained with a few data sets, and the training time required for deep learning is greatly shortened.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [28] Jiang Z. Dong Z. Jiang W. Yang Y. Recognition of rice leaf diseases and wheat leaf diseases based on multi-task deep transfer learning Computers and Electronics in Agriculture 2021 186 106184 10.1016/j.compag.2021.106184 source

      Not checked in the free run.

    • cites [29] Zhen-Cun J. Xiao-Jing W. E. N. Zheng-Xin D. Research on fire detection of improved VGG16 image recognition based on deep learning Fire Science and Technology 2021 40 3 p. 375

      Not checked in the free run.

    • cites [30] Wen-Ping J. Zhen-Cun J. Research on early fire detection of Yolo V5 based on multiple transfer learning Fire Science and Technology 2021 40 1 109 112

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisereasoning

    The vision transformer model and the CNN model extract features in completely different ways.

    “The vision transformer model is an image classification model based entirely on the transformer structure, which has completely different feature extraction method from the CNN model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The RAdam optimizer has good performance in convergence effect and convergence speed and can stably improve accuracy compared with SGD and Adam optimizers.

    “RAdam optimizer has good performance in convergence effect and convergence speed and is relatively compared with Adam optimizer, and it can stably improve accuracy.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Liu L. Jiang H. He P. On the variance of the adaptive learning rate and beyond 2019 http://arxiv.org/abs/1908.03265

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The ViT-CNN ensemble model is the first time the vision transformer model and the CNN model have been combined to diagnose acute lymphocytic leukemia.

    “This is the first time the vision transformer model and the CNN model have been combined to diagnose acute lymphocytic leukemia.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The RAdam optimizer is based on the Adam optimizer and introduces a correction to the adaptive learning rate to correct the variance of the Adam optimizer.

    “The RAdam optimizer is based on the Adam optimizer, which aims to introduce a correction to the adaptive learning rate to correct the variance of the Adam optimizer [32].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Liu L. Jiang H. He P. On the variance of the adaptive learning rate and beyond 2019 http://arxiv.org/abs/1908.03265

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The vision transformer model is completely implemented based on the transformer structure without any CNN structure.

    “The vision transformer model is completely implemented based on the transformer structure without any CNN structure [22].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Dosovitskiy A. Beyer L. Kolesnikov A. An image is worth 16 × 16 words: transformers for image recognition at scale 2020 http://arxiv.org/abs/2010.11929

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Fine-tuning increases the convergence speed and generalization ability of the model and reduces the risk of overfitting.

    “Fine-tuning not only increases the convergence speed and generalization ability of the model but also reduces the risk of overfitting [28–30].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [28] Jiang Z. Dong Z. Jiang W. Yang Y. Recognition of rice leaf diseases and wheat leaf diseases based on multi-task deep transfer learning Computers and Electronics in Agriculture 2021 186 106184 10.1016/j.compag.2021.106184 source

      Not checked in the free run.

    • cites [29] Zhen-Cun J. Xiao-Jing W. E. N. Zheng-Xin D. Research on fire detection of improved VGG16 image recognition based on deep learning Fire Science and Technology 2021 40 3 p. 375

      Not checked in the free run.

    • cites [30] Wen-Ping J. Zhen-Cun J. Research on early fire detection of Yolo V5 based on multiple transfer learning Fire Science and Technology 2021 40 1 109 112

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Literature [11] has the best current research results on diagnosis of acute lymphoblastic leukemia.

    “Literature [11] has the best current research results on diagnosis of acute lymphoblastic leukemia.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [11] Kasani P. H. Park S.-W. Jang J.-W. An aggregated-based deep learning method for leukemic B-lymphoblast classification Diagnostics 2020 10 12 p. 1064 10.3390/diagnostics10121064 PMC7763941 33302591 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The ViT-CNN ensemble model can accurately distinguish between cancer cells and normal cells and can be used as an effective method for computer-aided diagnosis of acute lymphoblastic leukemia.

    “The proposed method can accurately distinguish between cancer cells and normal cells and can be used as an effective method for computer-aided diagnosis of acute lymphoblastic leukemia.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Interpretationown data

    The ViT-CNN ensemble model has a better ability to classify cancer cells and normal cells than Resnet50, Densenet121, and VGG16.

    “This shows that the ViT-CNN ensemble model has a better ability to classify cancer cells and normal cells.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The ViT-CNN ensemble model has better classification performance than other models and can assist in the diagnosis of acute lymphocytic leukemia more accurately.

    “Obviously, the ViT-CNN ensemble model has better classification performance and can assist in the diagnosis of acute lymphocytic leukemia more accurately.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The symmetric cross-entropy loss function reduces the influence of noise and prevents overfitting in medical image datasets.

    “The symmetric cross-entropy loss function can reduce the influence of noise and prevent overfitting [31].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [31] Wang Y. Ma X. Chen Z. Symmetric cross entropy for robust learning with noisy labels Proceedings of the IEEE/CVF International Conference on Computer Vision 2019 Montreal, Canada 322 330 10.1109/iccv.2019.00041 source

      Not checked in the free run.

    • Rests on an assumption: Medical image datasets contain unavoidable noise, and the morphological similarity of cancer and normal cells results in some label noise.
20 of 20

  • The paper relies on this without support.

    Directly training on an unbalanced data set can easily cause the model to fall into overfitting or cause the generalization ability of the model to be weak.

    “Directly training on an unbalanced data set can easily cause the model to fall into overfitting or cause the generalization ability of the model to be weak.”

  • The paper relies on this without support.

    Medical image datasets contain unavoidable noise, and the morphological similarity of cancer and normal cells results in some label noise.

    “For medical images data sets, a bit of noise is unavoidable.”
    “The morphological similarity of cancer cells and normal cells results in some noise on the label.”

  • The paper relies on this without support.

    The greater the differences between the models, the greater the performance improvement after the ensemble.

    “The greater the differences between the models, the greater the performance improvement after the ensemble.”
3 of 3

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Most works in medical image classification jointly adopt CNN and transformer but did not develop a transformer-based backbone without convolutions to capture multi-granularity information in medical images.

    “Most works [30 -34] in medical image classi fi cation jointly adopt the CNN and transformer model but did not develop a transformer-based backbone without convolutions to capture multi-granularity information in medical images.”

    From DilatedFormer: dilated granularity transformer network for placental maturity grading in ultrasound (2023) · cites this paper as [32]

    Section 2.6 describes this medical-image classifier as an ensemble of a vision transformer and a CNN, so the paper is an example of jointly using the two models. Section 2.5.1 says its vision-transformer component has no CNN structure. The paper does not establish that *most* works use this approach or discuss developing a convolution-free backbone to capture multi-granularity information. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

The citation lists haven’t been fetched yet

References 27 · Cited by 124

Research works printed but not linked: 4

  1. Peng X., 2019, Application of cell morphology-related technology in hematological tumors, Chinese Journal of Biological Engineering, 39, 84
  2. DuggalR. GuptaA. andGuptaR. Segmentation of overlapping/touching white blood cell nuclei using artificial neural networks Proceedings of the CME Series on Hemato-Oncopathology July 2016 New Delhi India All India Institute of Medical Sciences (AIIMS).
  3. TanM.andLeQ. Efficientnet: rethinking model scaling for convolutional neural networks Proceedings of the International Conference on Machine Learning June 2019 Long Beach CA USA 6105–6114.
  4. PaszkeA. GrossS. MassaF.et al. Pytorch: an imperative style high-performance deep learning library 2019 http://arxiv.org/abs/912.01703.

Authors
Jiang, Z., Dong, Z., Wang, L., & Jiang, W.
Affiliations
  • Shanghai Institute of Technology (CN)
  • Shanghai Jiao Tong University (CN)
Corresponding author
Jiang, Wenping
ORCID
Venue
Computational Intelligence and Neuroscience
Publisher
Wiley
Published
2021-01
Volume
2021
Issue
1
Pages
7529893-7529893
ISSN
  • 1687-5265 (ISSN-L)
  • 1687-5273
DOI
10.1155/2021/7529893
PMC
PMC8405335
OpenAlex
W3194502875
Cited by
124
References
27
Citations by year
2026: 17 · 2025: 32 · 2024: 30 · 2023: 30 · 2022: 14 · 2021: 1
FWCI
6.74
Type
journal article
Language
en
Open access
Licence
Locations
  • Computational Intelligence and Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Natural Science Foundation of China (81827807, 61675134, 61307015)
  • Science and Technology Commission of Shanghai Municipality (19441905800)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Imaging for Blood Diseases
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Cell Image Analysis Techniques
Keywords
acute lymphoblastic leukemia, computer-aided diagnosis, cell image classification, vision transformer, convolutional neural networks, ensemble learning, class imbalance, noise robustness
MeSH terms
Diagnosis, Computer-Assisted, Humans, Neural Networks, Computer, Precursor Cell Lymphoblastic Leukemia-Lymphoma

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review