henhao
Sign in
BioinformaticsbioRxiv
Preprint

Breast Cancer Histopathological Image Classification: A Deep Learning Approach

Mehdi Habibzadeh Motlagh, Mahboobeh Jannesari, Hamidreza Aboulkheyr EsiD, Pegah KhosraviiD, Olivier ElementoiD, Mehdi TotonchiiD, Iman HajirasoulihaiD

Abstract

Breast cancer remains the most common type of cancer and the leading cause of cancer-induced mortality among women with 2.4 million new cases diagnosed and 523,000 deaths per year. Historically, a diagnosis has been initially performed using clinical screening followed by histopathological analysis. Automated classification of cancers using histopathological images is a chciteallenging task of accurate detection of tumor sub-types. This process could be facilitated by machine learning approaches, which may be more reliable and economical compared to conventional methods.\n\nTo prove this principle, we applied fine-tuned pre-trained deep neural networks. To test the approach we first classify different cancer types using 6, 402 tissue micro-arrays (TMAs) training samples. Our framework accurately detected on average 99.8% of the four cancer types including breast, bladder, lung and lymphoma using the ResNet V1 50 pre-trained model. Then, for classification of breast cancer sub-types this approach was applied to 7,909 images from the BreakHis database. In the next step, ResNet V1 152 classified benign and malignant breast cancers with an accuracy of 98.7%. In addition, ResNet V1 50 and ResNet V1 152 categorized either benign- (adenosis, fibroadenoma, phyllodes tumor, and tubular adenoma) or malignant- (ductal carcinoma, lobular carcinoma, mucinous carcinoma, and papillary carcinoma) sub-types with 94.8% and 96.4% accuracy, respectively. The confusion matrices revealed high sensitivity values of 1, 0.995 and 0.993 for cancer types, as well as malignant- and benign sub-types respectively. The areas under the curve (AUC) scores were 0.996,0.973 and 0.996 for cancer types, malignant and benign sub-types, respectively. Overall, our results show negligible false negative (on average 3.7 samples) and false positive (on average 2 samples) results among different models. Availability: Source codes, guidelines and data sets are temporarily available on google drive upon request before moving to a permanent GitHub repository.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
125
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (4 Jan 2018)

Fine-tuned ImageNet ResNets beat Inception models on tissue micro-array and BreaKHis slides, reaching 99.8% across four tumour tissues and 96.4% across four malignant sub-types.

Manual histopathological grading is slow and error-prone, and earlier automated work mostly stopped at a benign-versus-malignant call, used hand-crafted features, or trained a separate classifier per magnification. Here ImageNet-pretrained Inception and ResNet networks are fine-tuned to separate four cancer tissues and, within breast tissue, four benign and four malignant sub-types from images pooled across magnifications.

Key results

  1. ShownOn tissue micro-array images of four tumour tissues (breast, lung, bladder, lymphoma), ResNet variants classified nearly all samples correctly and clearly outperformed every Inception architecture tested.

    ResNet V1 50 with all layers fine-tuned reached 99.8%, ResNet V1 101/152 99.6%, and ResNet accuracies spanned 98.3-99.8%, against 76.4-97.1% for Inception (best: Inception V1 at 97.1%, 3,000 epochs). Cohen's unweighted kappa was above 0.97 for ResNet versus 0.94, 0.94, 0.82 and 0.84 for Inception V1-V4; mean false positives 0.3 versus 82. · A. Classification of Cancer Types

  2. ShownSeparating benign from malignant breast images in BreaKHis was also more accurate with fully fine-tuned ResNets than with any Inception model.

    ResNet V1 101 with all layers fine-tuned reached 98.4% on a 90%/10% split, versus 97.8% for ResNet V1 50 (all layers) and 94.1% for ResNet V1 101 with only the last layer; Inception V2 topped the Inception models at 94.1%. Mean false positives were 6.3 for ResNet against 15.25 for Inception. The abstract instead reports 98.7% for ResNet V1 152 on this task. · B. Malignant and Benign Breast Cancer; Abstract

  3. ShownWithin each class, the networks also separated the four malignant sub-types (ductal, lobular, mucinous, papillary carcinoma) and the four benign sub-types, with accuracy rising as training ran longer.

    Malignant sub-types on the test set: 96.4% for ResNet V1 152 and 94.6% for ResNet V1 50 with all layers fine-tuned, 90% with last-layer fine-tuning only, and 86.6% for Inception V1 (all layers, 3,000 epochs); mean false positives 0.33 for ResNet versus 0.75 for Inception. Longer training raised ResNet 101 to 96.5% and 98.5%. The abstract gives 94.8% for benign sub-types. · C. Breast Cancer Sub-types Classification; Abstract

  4. SuggestedReported discrimination was near-ceiling on all three tasks, with few missed positives, but these summary metrics appear only in the abstract and conclusion, not with per-model detail in the text.

    Sensitivity 1, 0.995 and 0.993 and AUC 0.996, 0.973 and 0.996 for cancer types, malignant sub-types and benign sub-types; on average 3.7 false negatives and 2 false positives across models. The conclusion quotes false positive counts of 0.3 out of 900 for cancer types, 6.3 out of 809 for breast cancer, 5 out of 800 benign and 0.3 out of 1000 malignant. · Abstract; V. Discussion

  5. ClaimedThe authors place these numbers above published results for hand-crafted-feature classifiers and earlier deep models, but no competing method was re-run here.

    The discussion cites 90 to 93% for conventional machine learning with hand-crafted features (SVM, KNN, QDA and variants) evaluated per magnification, 90% for an AlexNet benign-versus-malignant model, and 93.2% average accuracy for a published multi-class deep learning approach; none of these was reproduced on the same split. · V. Discussion

Methods, data and reproducibility

H&E and immunohistochemistry images were converted to RGB, augmented (resize, rotate, crop, flip), normalised and resized to 299x299x3 or 224x224x3, then classified with ImageNet-pretrained Inception (V1-V4) and ResNet (V1 50, 101, 152) networks fine-tuned either at the last layer only or throughout, using RMSProp, dropout and batch normalisation.

  • Tissue Micro Array (TMA) histopathological images of lung, breast, lymphoma and bladder cancer tissue · human · n = 6,402 images
  • BreaKHis breast images pooled across 40X, 100X, 200X and 400X magnification; four benign sub-types (adenosis, fibroadenoma, phyllodes tumor, tubular adenoma) and four malignant carcinoma sub-types · human · n = 7,909 images (2,480 benign, 5,429 malignant) from 82 patients
Methods & tools
TensorFlow (TF-slim), Inception V1-V4, ResNet V1 50/101/152, Augmentor, ImageNet pretrained weights, RMSProp
Reagents & environment
CentOS 7.2 64-bit with Python 3.5.3, GPU-enabled TensorFlow with CUDA 8.0 Toolkit and cuDNN v5.1.
Assay conditions & compute
A single HP DL380 G9 server: two E5-2690v4 CPUs (2.6 GHz, 14C), 64 GB DDR4 RAM and one ASUS GeForce GTX 1080 (2560 CUDA cores, 8GB GDDR5); batch size 32 in training and 100 in evaluation. No training time is given.
Reproducibility gaps
Code, guidelines and data are described as temporarily on Google Drive on request, pending a future repository; no URL appears in the extracted text, and the per-model tables and confusion matrices are in supplementary files not included.
Limitations (4)
  • Training sets were small relative to the model parameter counts, which the authors address with ImageNet pretraining, augmentation, dropout and batch normalisation rather than more data. · stated by the authors
  • Only classification was attempted; the authors suggest segmentation architectures such as U-net and DeepLab v3 for future work. · stated by the authors
  • Images were split randomly rather than by patient, and BreaKHis holds multiple magnifications of the same lesions from 82 patients, so near-duplicate views of one patient can fall on both sides of the split and inflate accuracy. · noted in this summary
  • Numbers disagree between sections: the split is 85%/15% in the methods but 90%/10% in the results, and the abstract's benign-versus-malignant result (98.7%, ResNet V1 152) differs from the results text (98.4%, ResNet V1 101); no repeated runs, confidence intervals or external test set are reported. · noted in this summary

Who should read:Read this if you are choosing a pretrained backbone and fine-tuning depth for histopathology image classification, and want ResNet-versus-Inception comparisons on tissue micro-array and BreaKHis data.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISRakhlin '18Phan '20Gella '24Khosravi '17Hungund '25Mujawar '25
80% semantic matchPathology
Alexander Rakhlin, Alexey Shvets, Vladimir Iglovikov+1 more·bioRxiv·Apr 2, 2018