A Novel method for IDC Prediction in Breast Cancer Histopathology images using Deep Residual Neural Networks

Chandra Churh Chatterjee; Gopal Krishna

REVIEW 4 major objections 7 minor 21 references

A Novel method for IDC Prediction in Breast Cancer Histopathology images using Deep Residual Neural Networks

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Seven-channel residual network learns IDC patches at 99.29% accuracy

desk verdict A routine patch-level CNN paper with an untrustworthy evaluation protocol; the 99.29% accuracy is likely inflated by slide-level leakage and cannot be taken at face value. read the letter →

arxiv 1908.07362 v2 pith:HCQEZ344 submitted 2019-08-20 eess.IV cs.CV

Chandra Churh Chatterjee , Gopal Krishna This is my paper · ORCID

classification eess.IVcs.CV

keywords invasiveductalcarcinomahistopathologyimageclassificationdeepresidualnetworkseven-channelcolorspaceCIELABHSVGrad-CAMbreastcancerdetection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a deep residual convolutional network can classify invasive ductal carcinoma (IDC) in breast histopathology patches almost perfectly. On a 7,500-image subset of the Breast Histopathology Images benchmark, the authors convert each $50 \times 50$ RGB patch into a seven-channel matrix—adding hue and saturation from HSV and $l^*$ and $a^*$ from CIELAB—then feed it through four residual blocks with $4 \times 4$ kernels. The reported result is 99.29% test accuracy with an AUROC of 0.9996, and the authors argue this beats previously published CNN and feature-based detectors. If the claim holds, automated IDC screening on whole-slide images becomes substantially more reliable, and the seven-channel color preprocessing is a reusable recipe for stain-heavy microscopy.

What carries the argument

The load-bearing mechanism is the seven-channel image matrix: RGB plus the hue and saturation channels of HSV plus the $l^*$ and $a^*$ channels of CIELAB, all denoised with a Gaussian blur and contrast-enhanced with CLAHE on the RGB channels only. The paper argues these extra color channels let the network see past stain variation. The network itself is a residual CNN with four residual blocks, each block stacking three convolutional layers and an unactivated shortcut connection whose outputs are added before ReLU; a global average pooling layer feeds a fully connected classifier.

What would settle it

A patient-level or slide-level split—placing all patches from each of the 162 source slides entirely in train, validation, or test—would settle it: if accuracy or AUROC falls materially below 99.29% and 0.9996, the original numbers reflected patch-level leakage rather than true IDC detection. An additional RGB-only run would isolate whether the seven-channel input is responsible for the gain.

Watch

Extended reading notes

Core claim

The central claim is that a 4-block deep residual network trained on a seven-channel image representation, rather than plain RGB, distinguishes IDC-positive from IDC-negative histopathology patches at essentially ceiling performance. The authors report per-class accuracy of 0.9929, precision and recall around 0.99 for both classes, and AUROC 0.9996, and they use Grad-CAM heatmaps to argue the network focuses on the affected regions. They frame this as a state-of-the-art result relative to the earlier CNN and machine-learning methods listed in their comparison table.

Load-bearing premise

The evaluation assumes randomly splitting patches into train and test does not leak information from the same whole-slide image into both sets; if that assumption fails, the reported near-perfect scores overstate how well the model would do on new patients.

Editorial extensions

If this is right

The 99.29% accuracy and 0.9996 AUROC, if taken at face value, mean the model misclassifies only about one in 140 test patches, far below the 81–89% accuracies of the compared methods.
The seven-channel preprocessing is, on the paper's account, what lets the network handle stain variation and low-resolution $50 \times 50$ patches.
Grad-CAM visualizations, as the authors use them, provide patch-level localization of the regions driving the IDC decision, which is useful for pathologist review.
The model's ceiling-level performance on this subset implies the remaining research question is generalization to larger and more varied histopathology cohorts, which the authors identify as future scope.

Reading between the lines

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

A patient-level or whole-slide-level split would be the real test of the 99.29% figure; because patches are drawn from only 162 source slides, random patch splitting likely places highly correlated neighbours in train and test, inflating the score.
An RGB-only ablation would quantify how much the four extra channels contribute; without it, the seven-channel claim is untested.
Applying the exact pipeline to the full 277,524-patch set or to an outside hospital's slides would show whether the near-perfect scores survive scale and stain shift.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The one thing to know: this is a standard residual CNN applied to a well-known Kaggle IDC patch dataset, with a small twist of feeding a 7-channel input (RGB + HSV hue/saturation + CIELAB L*/a*). The reported 99.29% accuracy and 0.9996 AUROC are not credible as estimates of generalization, because the evaluation split is at patch level, not patient or whole-slide level, and the dataset has only 162 source WSIs. The paper explicitly says only that no images overlap between train/test/validation; it never says patches from the same WSI were kept together. The stress-test note is correct, and it's load-bearing.

What's actually new: the 7-channel fusion is the only real novelty. No ablation isolates it. Everything else—residual blocks, ReLU/ELU, CLAHE, Gaussian blur, Grad-CAM—is off-the-shelf. The paper does cite the relevant prior work (He et al., Bejnordi et al., Reza & Ma), so the citation pattern is honest even if the comparisons in Table II are apples-to-oranges. The writing is clear enough, and the architecture description is reproducible in principle. The authors also disclose their subset selection (7,500 of 277,524 patches) and their hyperparameter tuning range, which is more than many similar preprints do.

Soft spots, in proportion: the single-split evaluation with no error bars or repeated runs is a serious but fixable flaw. The absence of any patient-level analysis or code/data release makes the headline numbers unverifiable. The state-of-the-art comparison mixes patch-level and WSI-level tasks; the Bejnordi stacked-CNN result is a three-class WSI accuracy, not directly comparable. The "one of a kind state-of-the-art-model" phrase is hype, but that's minor. The core problem remains the split: with 162 slides, the effective sample size is at most 162, and any patch-level accuracy can be inflated by slide-specific staining artifacts. That alone justifies rejection as it stands.

Who this is for: someone doing a quick literature sweep on IDC patch classification might skim it for the 7-channel idea, but the result should not be cited as evidence for a clinically useful system. It deserves a serious referee only if revised with a slide-level split, ablations, and confidence intervals. The current version would not survive peer review. My recommendation: desk reject now, but tell the authors what a revised version would need. It's not a fundamentally broken idea—it's an inadequately evaluated one.

Referee Report

4 major / 7 minor

Summary. The paper proposes a deep residual convolutional neural network for classifying invasive ductal carcinoma (IDC) in breast histopathology images. The authors take a 7,500-patch subset of the Kaggle Breast Histopathology Images dataset (originally 277,524 patches from 162 whole-slide images), preprocess the patches by resizing to 100×100, merging RGB with hue/saturation from HSV and l*/a* from CIELAB into a 7-channel input, applying Gaussian blur and CLAHE, and then train a network with four residual blocks and 4×4 kernels. The paper reports 99.29% accuracy and an AUROC of 0.9996 on a held-out test set, presents a confusion matrix and Grad-CAM visualizations, and claims state-of-the-art performance compared with earlier methods.

Significance. If the reported performance were obtained under a valid evaluation protocol, the method would represent a strong result for patch-level IDC classification and might be of practical interest for computer-aided diagnosis. The use of a multi-channel color representation, residual learning, and Grad-CAM visualization are sensible ingredients. However, the significance as presented is severely limited by the evaluation: the effective number of independent units is at most 162 whole-slide images, and the paper does not demonstrate that the train/test split is disjoint at the slide or patient level. Without such a split, the reported 99.29% accuracy and 0.9996 AUROC cannot be interpreted as estimates of generalization to new slides or patients, and the state-of-the-art claim is therefore not supported.

major comments (4)

[Section IV-A] The evaluation protocol is the load-bearing issue. The dataset consists of 277,524 patches extracted from 162 whole-slide images, and the authors use a 7,500-patch subset. Section IV-A states only that "No images overlap with each other out of the training, testing and the validation set of images," which is compatible with random patch-level splitting. If patches from the same whole-slide image appear in both training and test sets, the model can exploit slide-specific staining and preparation artifacts, inflating the reported 99.29% accuracy and 0.9996 AUROC. The authors must either perform a slide-level or patient-level split, or otherwise explicitly demonstrate that no two patches from the same source WSI are distributed across different splits, before the headline numbers can be accepted as valid performance estimates.
[Section IV-A and Table I] All performance metrics are reported for a single random split, with no repeated runs, cross-validation, or confidence intervals. Given that the independent units are at most 162 whole-slide images, the point estimates in Table I (e.g., 0.9929 accuracy, 0.9996 AUROC) are not accompanied by any measure of variability. The authors should report mean and standard deviation over multiple slide-disjoint splits or use slide-level cross-validation to establish that the results are stable.
[Table II and Section IV-C] The comparative analysis in Table II is not valid as presented. The cited prior works use different datasets, different patch-sampling strategies, different class balances, and different evaluation protocols. The claim that the proposed model "outperforms" these methods and is "an one of a kind state-of-the-art-model" cannot be sustained without a common benchmark and identical evaluation conditions. At minimum, the authors should clearly state the incomparability and restrict their claims to their own dataset and protocol.
[Section III-A and Section IV] The process for selecting the 7,500-image subset is not described: it is not stated whether the subset was chosen randomly, in a stratified manner, or consecutively from particular whole-slide images. Additionally, the hyperparameter search (kernel dimension and number of residual blocks) is reported only as a final choice, and the authors state that models with 1–3 and 5 residual blocks "produced poor results" without reporting the actual values. This makes it impossible to assess overfitting or selection bias. The authors should describe the subset selection and report the full hyperparameter sweep, ideally on a validation set that is separate from the test set.

minor comments (7)

[Abstract and Section IV-C] The phrase "an one of a kind state-of-the-art-model" should be corrected to "a one-of-a-kind state-of-the-art model."
[Section III-A] There is a typo: "neagtive" should be "negative."
[Section III-C] "Kaming He" should be "Kaiming He" in the reference to the residual network paper.
[Section IV-B] The phrase "over the top accuracy" is informal; please use a neutral formulation such as "very high accuracy" or simply report the numerical value.
[Section IV-C] Figure 8 is referenced before Figure 6 in the text; please renumber or reorder the figures so that they are cited in numerical order.
[Section III-C] The architecture description does not specify the number of convolutional filters per layer, the stride, or the input channel dimension after the 7-channel merge. These details are necessary for reproducibility.
[References] Reference [1] is cited in the abstract as the American Cancer Society, but the actual reference is to breastcancer.org; please align the citation with the intended source.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy and AUROC are empirical test-set measurements, not derivations from fitted inputs or self-cited premises.

full rationale

The paper's central claim is an empirical classification performance on a held-out test split of the Breast Histopathology Images benchmark. The 99.29% accuracy and 0.9996 AUROC come from a confusion matrix and ROC curve computed on test predictions, as stated in Section IV. The model is a residual CNN with a 7-channel input formed from RGB, HSV, and CIELAB channels; these are preprocessing choices, not parameters fitted to the test labels. The authors tune kernel sizes and residual-block counts using training/validation performance, which is standard model selection rather than circular reasoning. There are no self-citations to prior work by the same authors, no invoked uniqueness theorem, and no renamed empirical result presented as derivation. The most serious concern is that the train/test split is at the patch level rather than the whole-slide or patient level, so patches from the same source whole-slide image may appear in both training and test sets, potentially inflating the reported numbers through slide-level confounders. That is an evaluation-validity or soundness problem, not circularity: the prediction is not equivalent to the input by construction, and no fitted parameter is relabeled as a prediction. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a specific data subset, hand-selected hyperparameters, a 7-channel representation without ablation, and an assumed valid random split. The network weights themselves are fitted to the training data and are not released, so the reported accuracy cannot be verified independently.

free parameters (5)

CNN weights = Learned from 6,000 training patches
All reported metrics come from a single trained model; weights are not released, so the fitted values cannot be inspected or reproduced.
Training/test subset composition = 7,500 images (3,000 IDC-positive, 4,500 negative) sampled from 277,524 patches
The subset selection is described only as a response to memory errors (Section III-A); the sampling method is not stated, and the class balance differs from the full dataset.
Residual block count = 4
Chosen by fine-tuning over 1 to 5 blocks; configurations other than 4 'produced poor results' and were excluded (Section IV).
Kernel dimension = 4x4
Chosen by fine-tuning from 2x2 to 7x7; 4x4 gave the reported optimal curves (Section IV-C).
7-channel color representation = RGB plus HSV H,S and CIELAB L*,a*
Proposed preprocessing choice (Section III-B); no ablation is provided to show it improves over RGB alone.

assumptions (3)

domain assumption The Kaggle dataset ground-truth labels for IDC are correct as provided.
The paper uses the benchmark labels without independent pathological review or label-quality analysis (Section III-A).
domain assumption Patches from the same whole-slide image are independent and can be split randomly into train and test.
The train/test protocol in Section IV-A only guarantees no pixel-level image overlap; it does not separate patients or source slides, so the assumed independence is likely violated.
domain assumption Gaussian blur and CLAHE preprocessing preserve the diagnostic signal needed for classification.
The preprocessing choices in Section III-B are applied without an ablation or sensitivity analysis showing they help rather than discard information.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Novel method for IDC Prediction in Breast Cancer Histopathology images using Deep Residual Neural Networks." pith.science (2026). https://pith.science/paper/HCQEZ344

@misc{pith2026190807362,
  author       = {Pith},
  title        = {Pith review of: A Novel method for IDC Prediction in Breast Cancer Histopathology images using Deep Residual Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCQEZ344}},
  note         = {Machine review of arXiv:1908.07362}
}

read the original abstract

Invasive ductal carcinoma (IDC), which is also sometimes known as the infiltrating ductal carcinoma, is the most regular form of breast cancer. It accounts for about 80% of all breast cancers. According to the American Cancer Society, more than 180,000 women in the United States are diagnosed with invasive breast cancer each year. The survival rate associated with this form of cancer is about 77% to 93% depending on the stage at which they are being diagnosed. The invasiveness and the frequency of the occurrence of these disease makes it one of the difficult cancers to be diagnosed. Our proposed methodology involves diagnosing the invasive ductal carcinoma with a deep residual convolution network to classify the IDC affected histopathological images from the normal images. The dataset for the purpose used is a benchmark dataset known as the Breast Histopathology Images. The microscopic RGB images are converted into a seven channel image matrix, which is then fed to the network. The proposed model produces a 99.29% accurate approach towards the prediction of IDC in the histopathology images with an AUROC score of 0.9996. Classification ability of the model is tested using standard performance metrics.

Figures

Figures reproduced from arXiv: 1908.07362 by the authors.

**Figure 2.** Overview of the proposed methodology ductal carcinoma in situ (DCIS). B. E. Bejnordi et al. [16] put forward an algorithm to automatically detect DCIS in digitized hematoxylin and eosin (H&E) stained histopathological slides of breast tissue. This method utilizes multi-scale superpixel classification on whole slide images (WSIs) to distinguish DCIS from a large set of benign conditions. In [17] B. E. Bejnordi et al.… view at source ↗

**Figure 3.** Model architecture loss(caused by the problem of vanishinggradient) depending on the depth of the model. Hence our proposed methodology involves paradigms to overcome such problems [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: A single Residual Module from the proposed network [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Confusion Matrix of the proposed network [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: The comparison of the training accuracy and training loss curves for different kernel dimensions for the network with 4 residual blocks [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: The Grad-CAM visualization of the images(Left: Original processed images; Right: Grad-CAM heat map of the ROI of the images) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[1]

Invasive ductal carcinoma (idc),

Breastcancer.org, “Invasive ductal carcinoma (idc),” Accessed on: 2019, https://www.breastcancer.org/symptoms/types/idc

work page 2019
[2]

Invasive ductal carcinoma dataset,

Kaggle, “Invasive ductal carcinoma dataset,” Accessed on: 2019, https://www.kaggle.com/paultimothymooney/ breast-histopathology-images

work page 2019
[3]

Invasive ductal carcinoma treatment,

Cancer Treatement Centers of America, “Invasive ductal carcinoma treatment,” Accessed on: 2019, https://www.cancercenter.com/cancer-types/breast-cancer/risk-factors

work page 2019
[4]

Causes of breast cancer,

National Breast Cancer Foundation, INC, “Causes of breast cancer,” Accessed on: 2019, https://www.nationalbreastcancer.org/causes-of-breast-cancer

work page 2019
[5]

Gradient-based learning applied to document recognition,

Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

1998
[6]

Imagenet classiﬁcation with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classiﬁcation with deep convolutional neural networks,” in Advances in Neural Infor- mation Processing Systems 25 , F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds., 2012, pp. 1097–1105

work page 2012
[7]

Faster r-cnn: Towards real-time object detection with region proposal networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” in Advances in neural information processing systems , 2015, pp. 91–99

2015
[8]

Mask r-cnn,

K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask r-cnn,” in Computer Vision (ICCV), 2017 IEEE International Conference on . IEEE, 2017, pp. 2980–2988

work page 2017

Show all 21 references

[9]

Digital analysis of microscopic images in medicine,

N. Dey, A. S. Ashour, A. Ashour, and A. Singh, “Digital analysis of microscopic images in medicine,” Journal of Advanced Microscopy Research, vol. 10, pp. 1–13, 06 2015

2015
[10]

Deep residual learning for image recognition,

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

2016
[11]

A survey of data mining and deep learning in bioinformatics,

K. Lan, D.-t. Wang, S. Fong, L.-s. Liu, K. K. L. Wong, and N. Dey, “A survey of data mining and deep learning in bioinformatics,” Journal of Medical Systems , vol. 42, no. 8, p. 139, Jun 2018. [Online]. Available: https://doi.org/10.1007/s10916-018-1003-9

2018 doi
[12]

Breast cancer histopathology image analysis: A review,

M. Veta, J. P. W. Pluim, P. J. van Diest, and M. A. Viergever, “Breast cancer histopathology image analysis: A review,” IEEE Transactions on Biomedical Engineering , vol. 61, no. 5, pp. 1400–1411, 2014

2014
[13]

Automated grading of breast cancer histopathology using spectral clustering with textural and architectural image features,

S. Doyle, S. Agner, A. Madabhushi, M. Feldman, and J. Tomaszewski, “Automated grading of breast cancer histopathology using spectral clustering with textural and architectural image features,” 2008, pp. 496– 499

2008
[14]

Automated gland and nuclei segmentation for grading of prostate and breast cancer histopathology,

S. Naik, S. Doyle, S. Agner, A. Madabhushi, M. Feldman, and J. Tomaszewski, “Automated gland and nuclei segmentation for grading of prostate and breast cancer histopathology,” 2008, pp. 284–287

2008
[15]

Computerized classiﬁcation of intraductal breast lesions using histopathological images,

M. M. Dundar, S. Badve, G. Bilgin, V . Raykar, R. Jain, O. Sertel, and M. N. Gurcan, “Computerized classiﬁcation of intraductal breast lesions using histopathological images,” IEEE Transactions on Biomedical Engineering, vol. 58, no. 7, pp. 1977–1984, 2011

1977
[16]

Automated detection of dcis in whole-slide h & e stained breast histopathology images,

B. Ehteshami Bejnordi, M. Balkenhol, G. Litjens, R. Holland, P. Bult, N. Karssemeijer, and J. A. W. M. van der Laak, “Automated detection of dcis in whole-slide h & e stained breast histopathology images,” IEEE Transactions on Medical Imaging , vol. 35, no. 9, pp. 2141–2150, 2016

2016
[17]

Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images,

B. Ehteshami Bejnordi, J. Lin, B. Glass, M. Mullooly, G. L. Gierach, M. E. Sherman, N. Karssemeijer, J. van der Laak, and A. H. Beck, “Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images,” 2017, pp. 929–932

2017
[18]

Imbalanced histopathological breast cancer image classiﬁcation with convolutional neural network,

M. S. Reza and J. Ma, “Imbalanced histopathological breast cancer image classiﬁcation with convolutional neural network,” in 2018 14th IEEE International Conference on Signal Processing (ICSP) , 2018, pp. 619–624

2018
[19]

Invasive ductal breast carcinoma detector that is robust to image magniﬁcation in whole digital slides,

M. Balazsi, P. Blanco, P. Zoroquiain, M. Levine, and M. Burnier Jr, “Invasive ductal breast carcinoma detector that is robust to image magniﬁcation in whole digital slides,” Journal of Medical Imaging , vol. 3, p. 027501, 05 2016

2016
[20]

Automatic detection of invasive ductal carcinoma in whole slide images with convolutional neural networks,

A. Cruz-Roa, A. Basavanhally, F. Gonzlez, H. Gilmore, M. Feldman, S. Ganesan, N. Shih, J. Tomaszewski, and A. Madabhushi, “Automatic detection of invasive ductal carcinoma in whole slide images with convolutional neural networks,” Progress in Biomedical Optics and Imaging - Pr...

2014
[21]

Context- aware stacked convolutional neural networks for classiﬁcation of breast carcinomas in whole-slide histopathology images,

B. E. Bejnordi, G. Zuidhof, M. Balkenhol, M. Hermsen, P. Bult, B. van Ginneken, N. Karssemeijer, G. Litjens, and J. van der Laak, “Context- aware stacked convolutional neural networks for classiﬁcation of breast carcinomas in whole-slide histopathology images,” Journal of medi...

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Invasive ductal carcinoma (idc),

Breastcancer.org, “Invasive ductal carcinoma (idc),” Accessed on: 2019, https://www.breastcancer.org/symptoms/types/idc

work page 2019

[2] [2]

Invasive ductal carcinoma dataset,

Kaggle, “Invasive ductal carcinoma dataset,” Accessed on: 2019, https://www.kaggle.com/paultimothymooney/ breast-histopathology-images

work page 2019

[3] [3]

Invasive ductal carcinoma treatment,

Cancer Treatement Centers of America, “Invasive ductal carcinoma treatment,” Accessed on: 2019, https://www.cancercenter.com/cancer-types/breast-cancer/risk-factors

work page 2019

[4] [4]

Causes of breast cancer,

National Breast Cancer Foundation, INC, “Causes of breast cancer,” Accessed on: 2019, https://www.nationalbreastcancer.org/causes-of-breast-cancer

work page 2019

[5] [5]

Gradient-based learning applied to document recognition,

Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

1998

[6] [6]

Imagenet classiﬁcation with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classiﬁcation with deep convolutional neural networks,” in Advances in Neural Infor- mation Processing Systems 25 , F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds., 2012, pp. 1097–1105

work page 2012

[7] [7]

Faster r-cnn: Towards real-time object detection with region proposal networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” in Advances in neural information processing systems , 2015, pp. 91–99

2015

[8] [8]

Mask r-cnn,

K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask r-cnn,” in Computer Vision (ICCV), 2017 IEEE International Conference on . IEEE, 2017, pp. 2980–2988

work page 2017