Classification of molecular subtypes of breast cancer in whole-slide histopathological images using a deep learning algorithm

doi:10.21203/rs.3.rs-3240840/v1

Download PDF

Article

Classification of molecular subtypes of breast cancer in whole-slide histopathological images using a deep learning algorithm

https://doi.org/10.21203/rs.3.rs-3240840/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Classification of molecular subtypes of breast cancer is widely used in clinical decision-making, leading to different treatment responses and clinical outcomes. We classified molecular subtypes using a novel deep learning algorithm in whole-slide histopathological images (WSIs) with invasive ductal carcinoma of the breast. We obtained 1,094 breast cancer cases with available hematoxylin and eosin-stained WSIs from the TCGA database. We applied a new deep learning algorithm for artificial neural networks (ANNs) that is completely different from the back-propagation method developed in previous studies. Our model based on the ANN algorithm had an accuracy of 67.8% for all datasets (training and testing), and the area under the receiver operating characteristic curve was 0.819 when classifying molecular subtypes of breast cancer. In approximately 30% of cases, the molecular subtype did not reflect the unique histological subtype, which lowered the accuracy. The set revealed relatively high sensitivity (70.5%) and specificity (84.4%). Our approach involving this ANN model has favorable diagnostic performance for molecular classification of breast cancer based on WSIs and could provide reliable results for planning treatment strategies.

Breast cancer subtypes

Deep learning

Diagnostic performance

Histopathological images

Breast cancer is one of the most common cancers among women, with its prevalence increasing rapidly worldwide¹. Advancements in diagnostic methods and standardized treatments for breast cancer have reduced mortality; however, breast cancer is still the second leading cause of cancer mortality among women in the United States, and breast cancer mortality declines have slowed slightly in recent years². Breast cancer is a highly heterogeneous disease consisting of many different molecular subtypes with significant differences in prognosis and outcome. Identifying the molecular subtypes of breast cancer to address these heterogeneities is important for providing personalized treatments such as endocrine therapy, targeted therapy, and chemotherapy.

The development of RNA-based signature assays has made it possible to measure the expression of many genes in breast cancer, resulting in the classification of breast cancer into different molecular subtypes based on differences between tumors at the gene level^3–5. Molecular subtyping of breast cancer by RNA-based signature assays is increasingly being used as supplementary prognostic indicators because it has shown more significant differential survival than standard clinicopathologic factors⁶. Gene expression profiling is not routinely feasible. In clinical practice, breast cancer is commonly divided into four molecular subtypes using immunohistochemistry (IHC) surrogate markers, including estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), due to their cost-effectiveness, ease of implementation, and universal availability^3,7. The expression of these markers can guide the use of targeted therapies, such as hormonal therapy (for ER/PR-positive tumors) or HER2-targeted therapy (for HER2-positive tumors), and thus help predict response to treatment, disease progression, and survival^7–10. The development of more efficient methods to approximate the subtype classification of breast cancer based on RNA-based signature assays or IHC surrogate markers could be of great utility to accurately predict the prognosis and provide personalized treatment.

Machine learning methods have been gaining attention, including different deep learning algorithms, such as artificial neural networks (ANNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Deep learning can automatically learn by extracting features, such as important patterns and rules from data and making decisions or predictions based on information. Recent studies have shown the potential benefits of diagnostic performance by starting to apply deep learning using CNN to read or classify radiologic images^11–13. For pathological diagnosis, hematoxylin and eosin (H&E)-stained whole-slide histopathological images (WSIs) obtained by biopsy or surgical resection are subjected to detailed analysis by experienced pathologists. The pathological diagnosis suggests the types of cancer and information, such as the histological grade and subtype of malignant cells. Manual visual assessment is subject to interreader variability of pathologists for pathological diagnosis is error-prone, and time-consuming¹⁴. In addition, histopathologic findings that classify molecular subtypes are not internationally established, and in the absence of IHC support, histologic findings of an unestablished molecular subtype may lead to inappropriate treatment^15–17. In diagnostic pathology, these basic but important repetitive diagnostic tasks have the potential to be useful for automation via deep learning. It has recently become possible to use WSIs of H&E-stained glass slides through special scanning devices¹⁸. With the development of scanning technology, deep learning analysis has improved as histopathological images of various types of cancer have become available, but there have not been any studies evaluating the diagnostic performance of the ANN algorithm to classify molecular subtypes of breast cancer.

In this study, Pathologic slide images provide detailed information about the cell structure and shape of breast cancer tissues. By analyzing these images, deep learning algorithms can extract and quantify subtle visual features that are difficult for pathologists to identify. We assessed the classification performances of the ANN algorithm to identify four molecular subtypes (Luminal A, Luminal B, HER2 and Basal-like) in whole-slide images of breast cancer from The Cancer Genome Atlas (TCGA) database.

Classification of molecular subtypes using ANN models

We obtained 946 independent hematoxylin and eosin-stained histopathology whole-slide images of breast cancer from TCGA to evaluate the usefulness of our model. The cases were classified into four molecular subtypes: Luminal A (n = 616); Luminal B (n = 121); HER2-positive (n = 34); and Basal-like (n = 175). Table 1 shows the classification results of the molecular subtypes on the training and test sets using the ANN models. The results show the accuracy and error rate of the ANN model of 30 × 20, 36 × 24, 45 × 30, 60 × 40, 90 × 60 and 120 × 80 pixels separately for the training set and the test set, and the last row shows the average value of the accuracy and error rate for a given resolution. It was found that the highest accuracy and lowest error rate when diagnosing subtypes of breast cancer in the training set were at 36 × 24 pixels of 6 resolutions (0.752 and 0.389, respectively). We obtained the best diagnostic performance with the highest accuracy and lowest error rate when diagnosing subtypes of breast cancer in the test set validation (0.642 and 0.527, respectively) for the ANN structure including 9,600 input nodes (for the input pixel resolution of 120 × 80) and 40 intermediate nodes in the three hidden layers (Table 1).

Table 1

Classification results of the molecular subtypes on training and test sets using the artificial neural network.
Resolution (Pixel)	ANN Models		Training set		Test set
Resolution (Pixel)	Node	Hidden layer	Accuracy	Error rate	Accuracy	Error rate
30 \(\times\) 20	20	1	0.737	0.401	0.595	0.616
	40	1	0.728	0.405	0.605	0.600
	80	1	0.710	0.424	0.579	0.614
	20	2	0.742	0.397	0.574	0.648
	40	2	0.742	0.402	0.574	0.651
	80	2	0.724	0.419	0.574	0.632
	20	3	0.741	0.415	0.621	0.592
	40	3	0.733	0.406	0.579	0.627
	80	3	0.727	0.421	0.621	0.597
Average			0.732	0.410	0.591	0.620
36 \(\times\) 24	20	1	0.747	0.390	0.579	0.641
	40	1	0.734	0.400	0.600	0.627
	80	1	0.725	0.415	0.590	0.623
	20	2	0.752	0.389	0.579	0.655
	40	2	0.735	0.401	0.584	0.681
	80	2	0.731	0.408	0.574	0.639
	20	3	0.740	0.406	0.600	0.604
	40	3	0.751	0.399	0.579	0.650
	80	3	0.726	0.418	0.600	0.624
Average			0.738	0.403	0.587	0.638
45 \(\times\) 30	20	1	0.742	0.401	0.563	0.649
	40	1	0.726	0.412	0.595	0.610
	80	1	0.722	0.416	0.558	0.638
	20	2	0.751	0.392	0.526	0.707
	40	2	0.733	0.403	0.600	0.609
	80	2	0.727	0.418	0.600	0.622
	20	3	0.747	0.409	0.600	0.621
	40	3	0.744	0.401	0.574	0.672
	80	3	0.715	0.431	0.626	0.548
Average			0.734	0.409	0.582	0.631
60 \(\times\) 40	20	1	0.750	0.390	0.558	0.642
	40	1	0.737	0.404	0.579	0.614
	80	1	0.714	0.434	0.611	0.573
	20	2	0.744	0.406	0.579	0.628
	40	2	0.736	0.405	0.605	0.608
	80	2	0.706	0.430	0.568	0.632
	20	3	0.722	0.414	0.584	0.647
	40	3	0.724	0.422	0.553	0.637
	80	3	0.697	0.439	0.621	0.578
Average			0.726	0.416	0.584	0.618
90 \(\times\) 60	20	1	0.735	0.411	0.579	0.642
	40	1	0.724	0.433	0.584	0.572
	80	1	0.694	0.450	0.605	0.566
	20	2	0.734	0.417	0.568	0.620
	40	2	0.716	0.429	0.595	0.623
	80	2	0.685	0.456	0.590	0.571
	20	3	0.731	0.422	0.595	0.621
	40	3	0.715	0.439	0.595	0.591
	80	3	0.710	0.449	0.590	0.581
Average			0.716	0.434	0.589	0.599
120 \(\times\) 80	20	1	0.687	0.466	0.632	0.545
	40	1	0.660	0.496	0.642	0.539
	20	2	0.704	0.449	0.595	0.567
	40	2	0.671	0.487	0.642	0.544
	20	3	0.711	0.447	0.605	0.602
	40	3	0.678	0.476	0.642	0.527
Average			0.685	0.470	0.626	0.554
ANN, artificial neural network

Prediction accuracy of the 120 × 80 pixel model

Table 2 presents the diagnostic performance of our selected model (120 × 80 pixels) for classifying molecular subtypes on all datasets (training and testing). It was found that the model (node, 40; hidden layer, 3) had an accurate prediction rate of 67.8%, a sensitivity of 70.5%, a specificity of 84.4%, a positive predictive value (PPV) of 10.4%, and a negative predictive value (NPV) of 39.9% in predicting the molecular classification of breast cancer. The receiver operating characteristic curve (ROC) plots showed that the model (node, 40; hidden layer, 3) had an area under the curve (AUC) of 0.819 (95% CI, 0.782–0.856), with better discriminability than other models, such as “node, 40; hidden layer, 1” (AUC of 0.808, 95% CI, 0.770–0.845) and “node, 40; hidden layer, 2” (AUC of 0.807, 95% CI, 0.771–0.844) (Fig. 5).

Table 2

Diagnostic performance of deep-learning models for molecular classification of breast cancer on histopathological features in all datasets.
Resolution	Node	Hidden layer	Accuracy	Sensitivity	Specificity	PPV	NPV	AUC (95% CI)
120 \(\times\) 80	40	1	66.0%	68.4%	84.7%	11%	40.1%	0.808 (0.770–0.845)
		2	67.1%	65.3%	88.4%	11.6%	34.7%	0.807 (0.771–0.844)
		3	67.8%	70.5%	84.4%	10.4%	39.9%	0.819 (0.782–0.856)
PPV, positive predictive value; NPV, negative predictive value; AUC, area under the curve.

Recent studies have revealed that the deep learning approach has potential benefits in diagnostic performance for classifying and diagnosing medical images^11–13. We performed novel deep-learning algorithm models using an ANN approach to classify molecular subtypes of breast cancer in WSIs from the TCGA database^19,20. Our ANN algorithm with a resolution of 120 \(\times\) 80 pixels, 40 nodes and three hidden layers showed favorable diagnostic performance to classify molecular subtypes in all datasets: accuracy, 67.8%; AUC, 0.819; sensitivity, 70.5%; specificity, 84.4%. Previously, we predicted the location of the glottis in video laryngoscopy images and the detection and classification of intracranial hemorrhage on CT images and the detection of pneumoperitoneum in abdominal radiograph images through ANN with the “Kim-Monte Carlo algorithm”. The ANN model based on the “Kim-Monte Carlo algorithm” was developed by referring to the biological evolution of animals based on Monte Carlo simulation, which is different from the back-propagation method. Our previous studies based on the ANN algorithm demonstrated relatively high classification performance^19–21.

Predicting the molecular subtype of breast cancer plays a very important role in selecting the decision making about treatments and prognosis. Molecular subtyping of breast cancer by RNA-based signature assays is a well-established approach, but is not routinely performed due to cost-effectiveness, extended processing time, and need for appropriate tissue samples. To compensate for these limitations, IHC results, such as ER, PR and HER2, are widely used as references for planning breast cancer treatment. In some clinical practice, when IHC or genetic analysis cannot be performed due to poor sample stability and/or insufficient laboratory facilities, the subtype classification based on histopathological patterns is useful. Therefore, it would be a significant achievement if similar results could be obtained by replacing the classification of molecular subtypes based on gene expression with computer pattern recognition. It has recently become possible to use WSIs, digitized counterparts of glass slides obtained through special scanning devices. The subtype classification based solely on these histopathological images is available with more practical advice, and it is cost-effective and easier to access than expensive methods, such as molecular and/or IHC analyses. We attempted to apply the ANN algorithm to classify the subtypes on the basis of histopathological features of WSIs and showed relatively favorable results. Morphological patterns alone may not be sufficient for distinguishing molecular subtypes, but our analysis results could be used as a reference for planning molecular or histology studies.

Previous studies reported molecular subtype classification using breast radiologic or histologic images^22–24. A study reported that image data based on a CNN model were not sufficient for predicting the molecular subtype²⁴. Another study using a residual neural network (ResNet) was performed to classify the molecular subtypes in only 384 cases of breast cancer based on 4 pretrained models (VGG16, ResNet50, ResNet101, and Xception). The study included PAM50 molecular intrinsic subtypes and an accuracy of 68%²⁵. Another study using tissue microarray images (TMAs) reported that identification of ER-positive subtypes could be predicted with an accuracy of 75%²². The above studies used a small number of cases or used deep learning to distinguish a single subtype in limited image data, such as TMAs. Our study was the first attempt to classify the WSIs’ molecular subtypes using the ANN algorithm. Since our approach was not a CNN-based deep learning method, data selection and preprocessing of the input image to the ANN was not essential, and it has the advantage of using a simple training process and various small ANN models. Our novel deep‑learning algorithm did not show high enough accuracy to completely replace pathologist’s clinical diagnosis in the classification of molecular subtypes of breast cance. Previous studies have confirmed that WSI-based classification and PAM 50 subtyping may not be consistent because many of the expression-based features defined by the PAM50 gene are not visually identifiable features²³. Applying large-scale training datasets or applying data augmentation techniques such as image flipping, image magnification/reduction, and/or image translation to our training sets can significantly improve the diagnostic performance of our ANN models. Nevertheless, it did show favorable diagnostic performance, indicating that our deep learning-based tools and workflow system have the potential to support pathologists' prescreening and diagnostic double-reading of breast cancer subtypes to increase diagnostic accuracy and reduce effort.

Our results do not show more than 90% accuracy in classifying molecular subtypes. Several limitations of our study should be considered. First, some breast cancer cases do not have unique histopathological patterns that vary constantly according to the molecular subtypes. Second, there is a limit in that the present deep learning algorithms cannot learn minimal morphological changes according to the molecular subtypes. Third, in the WSIs, the proportion of supporting cells except for cancer cells may confuse the classification of molecular subtypes. Fourth, the resolution reduction of WSIs makes it impossible to recognize minimal histopathological changes according to the molecular subtypes. Fifth, there is an imbalance between breast cancer subtypes in the TCGA database used in the study. Sixth, the number of WSI data is insufficient to learn.

In summary, the novel deep-learning algorithm for ANNs has classified four molecular subtypes of breast cancer using WSIs from the TCGA database. Our approach is fully automated and skips preprocessing steps, resulting in faster processing times, and shows relatively good diagnostic performance and accuracy. The application of deep learning algorithms for subtyping breast cancer using pathology slide images can improve the accuracy, efficiency, and objectivity of the diagnostic process. Therefore, the clinical application of deep learning using WSI to improve the performance of these algorithms requires the development of additional programs and prospective validation studies using large datasets.

Case selection

In this study, we collected 1,094 breast cancer cases with H&E-stained WSI from the TCGA database (https://portal.gdc.cancer.gov/) to classify the molecular subtypes of breast cancer. This database is a publicly available database that includes samples from thousands of cancer patients²⁶. Clinicopathological parameters including mRNA expression values were obtained from “https://gdc.cancer.gov/about-data/publications/pancanatlas”²⁷. Out of the 1,094 images, 148 cases were excluded due to the lack of information on the molecular subtypes of breast cancer. We used 946 patients out of 1094 breast cancer patients for whom Predictor Analysis of Microarray 50 (PAM50) expression-based molecular subtyping results and WSIs were available in the TCGA database for this study. The cases were classified into four molecular subtypes: Luminal A (n = 616); Luminal B (n = 121); HER2-positive (n = 34); and Basal-like (n = 175)²⁸.

Image preprocessing

This study used the ANN algorithm approach. ANNs are a connected set of simple computing elements called ‘‘nodes.’’ They are usually organized into several layers: an input layer, an output layer, and several hidden layers in between²⁹. Each node in a hidden layer, i.e., an intermediate node, receives inputs from all the nodes in the previous layer, calculates a single output, and supplies it to all the nodes in the next layer. The signals are then propagated forward through the hidden layers to the output layer³⁰. The abstract representation of input data is extracted in the feed-forward process. The optimal number of hidden layers of an ANN depends on the character of the input data¹⁹(Fig. 1).

The slides were tiled in nonoverlapping several pixel windows at a magnification of 200× using Aperio’s ImageScope digital slide viewer (Aperio Technologies, Vista, CA). To obtain a fully connected ANN structure with optimal diagnostic performance, the input image resolution and the number of hidden layers in the ANN model started at 15 × 10 and 1, respectively, and were gradually increased to compare the resulting performances with those of other ANN models. The image resolutions were reduced to 120 × 80, 90 × 60, 60 × 40, 45 × 30, 36 × 24, and 30 × 20 pixels, and each resolution was applied to the input structure of an ANN model (Fig. 2). The number of input nodes of an ANN should be the total number of pixels of the square image with a reduced resolution. Thus, each black-and-white pixel value was divided by the maximum value of 255 to achieve a value between zero and 1.0 for use as the input value of an input node. The number of output nodes of an ANN should have been 4, i.e., 1 (Luminal A), 2 (Luminal B), 3 (HER2) or 4 (Basal-like) for classifying breast cancer into molecular subtypes. The number of hidden layers of an ANN was set to 1, 2, or 3. The number of intermediate nodes in a hidden layer was changed from 10 to 80. Accordingly, various ANN models, i.e., 99 models, were constructed.

Image augmentation was performed to generate more training images with the intention of avoiding overfitting and improving the overall performance of the models^31,32. We flipped (mirror) the images vertically or horizontally as follows: raw, vertical, horizontal and horizontal/vertical image (Fig. 3).

Training process

The deep-learning algorithm, “Kim-Monte Carlo algorithm,” a simple ANN training process that differed from the back-propagation method, was applied to train the various ANN models with the training set as learning data¹⁹. The algorithm applied a random optimization process based on a Monte Carlo simulation during the training session to determine a very large number of unknown weight factors and bias values for each ANN. All computer software programs for image processing and ANN were developed and programmed fully in-house by J. S. Kim using Microsoft Visual C + + and executed on the Windows 10 operating system of regular PCs without graphics processing unit (GPU) computing.

The initial weight factors and bias values of an ANN were randomly chosen within the ranges of -0.2 to + 0.2 and 0 to + 0.2, respectively. The randomly selected weight factors and bias values based on a given variable selection ratio were repeatedly adjusted by the amounts of the randomly picked delta-values within the range of − 0.1 to + 0.1, where the bias value was changed to positive if the variable value was changed to negative. Accepting or rejecting the adjustments depended on whether the new values decreased the average training error of the ANN for all training data, and 30 attempts were made to adjust the values of the selected variables in small random amounts. During a training cycle, the total sum of the variable selection ratio was set to 900% of the total number of variables in the ANN. A training session was terminated after 10 repetitions of the training cycle, during which the variable selection ratio of a training cycle was steadily decreased from 15–1.5%. After the ANN’s training session with the training set, the test set was applied to the ANN to obtain the test results, including the output values to identify the ANN model’s diagnostic performance.

For fully connected hidden nodes of an ANN model, a sigmoid activation function was applied and implemented by the following formulas:

\({y}_{j}=\frac{2}{1 +{ e}^{-{x}_{j}}} -1\)	(1)
\({y}_{j}=\frac{1}{1 +{ e}^{-{x}_{j}}}\)	(2)
\({x}_{j}={\sum }_{i}{w}_{ij}{y}_{i} + {b}_{j}\)	(3)

where Eq. (3) (using one of Equations (1) and (2)) denotes the summation over all nodes in the previous layer, i.e., each node j in a given layer receives an input \({y}_{i}\) from a node i in the previous layer; \({w}_{ij}\) indicates the weight factor between nodes i and j; and \({b}_{j}\) indicates the bias value of a node j. The sigmoid activation functions in Equations (1) and (2) were applied for fully connected hidden nodes and output nodes, respectively.

Figure 4 shows a computer screen displaying the ANN training progress. The gray curve denotes the ‘‘error rate’’, which is the average value of the training errors for all the training datasets, where the training error is the summation of all the output nodes of the square of the difference between the output value of an output node and the corresponding target value for given input data. Thus, the ‘‘error rate’’ is the absolute criterion of the training in progress. The horizontal value indicates the number of successfully changing values of the weight factors and the bias values of the ANN during the training session. The red curve denotes the ‘‘score’’, which is the average score for all the training datasets, where the score for a given input data is 1.0 if the output node with the maximum output value corresponds to the output node with the maximum target value and 0.0 for other cases. Thus, the ‘‘score’’ is a subsidiary reference value of the training in progress.

Statistical analysis

We performed ROC curve analysis to determine the diagnostic performances of various ANN algorithms for molecular subtype classification. Statistical analyses were conducted using commercially available software. All data were analyzed using R software packages and SPSS statistics (version 25.0, SPSS Inc., Chicago, IL, USA).

Data availability

The Data belongs to TCGA Research Network and is available in https://www.cancer.gov/tcga.

The authors declare that all data supporting the findings of this study are available within the article.

Acknowledgements

This research was supported by the Basic Science Research Program of the National Research Foundation funded by the Ministry of Education (RS-2022-00165754)

Competing interests

The authors declare no competing interests.

Siegel, R. L., Miller, K. D. & Jemal, A. Cancer statistics, 2019. CA Cancer J Clin 69, 7-34, doi:10.3322/caac.21551 (2019).
DeSantis, C. E. et al. Breast cancer statistics, 2019. CA Cancer J Clin 69, 438-451, doi:10.3322/caac.21583 (2019).
Perou, C. M. et al. Molecular portraits of human breast tumours. Nature 406, 747-752, doi:10.1038/35021093 (2000).
Sorlie, T. et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 98, 10869-10874, doi:10.1073/pnas.191367098 (2001).
van 't Veer, L. J. et al. Gene expression profiling predicts clinical outcome of breast cancer. Nature 415, 530-536, doi:10.1038/415530a (2002).
Nielsen, T. O. et al. A comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors in tamoxifen-treated estrogen receptor-positive breast cancer. Clin Cancer Res 16, 5222-5232, doi:10.1158/1078-0432.CCR-10-1282 (2010).
Goldhirsch, A. et al. Personalizing the treatment of women with early breast cancer: highlights of the St Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2013. Ann Oncol 24, 2206-2223, doi:10.1093/annonc/mdt303 (2013).
Andre, F. & Pusztai, L. Molecular classification of breast cancer: implications for selection of adjuvant chemotherapy. Nat Clin Pract Oncol 3, 621-632, doi:10.1038/ncponc0636 (2006).
Osborne, C. K. & Schiff, R. Mechanisms of endocrine resistance in breast cancer. Annu Rev Med 62, 233-247, doi:10.1146/annurev-med-070909-182917 (2011).
Pernas, S. & Tolaney, S. M. HER2-positive breast cancer: new therapeutic frontiers and overcoming resistance. Ther Adv Med Oncol 11, 1758835919833519, doi:10.1177/1758835919833519 (2019).
De Fauw, J. et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med 24, 1342-1350, doi:10.1038/s41591-018-0107-6 (2018).
Chilamkurthy, S. et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet 392, 2388-2396, doi:10.1016/S0140-6736(18)31645-3 (2018).
Ehteshami Bejnordi, B. et al. Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA 318, 2199-2210, doi:10.1001/jama.2017.14585 (2017).
Chung, Y. R. et al. Interobserver Variability of Ki-67 Measurement in Breast Cancer. J Pathol Transl Med 50, 129-137, doi:10.4132/jptm.2015.12.24 (2016).
Dobson, L. et al. Image analysis as an adjunct to manual HER-2 immunohistochemical review: a diagnostic tool to standardize interpretation. Histopathology 57, 27-38, doi:10.1111/j.1365-2559.2010.03577.x (2010).
Polley, M. Y. et al. An international Ki67 reproducibility study. J Natl Cancer Inst 105, 1897-1906, doi:10.1093/jnci/djt306 (2013).
Di Palma, S. et al. Chromogenic in situ hybridisation (CISH) should be an accepted method in the routine diagnostic evaluation of HER2 status in breast cancer. J Clin Pathol 60, 1067-1068, doi:10.1136/jcp.2006.043356 (2007).
Mukhopadhyay, S. et al. Whole Slide Imaging Versus Microscopy for Primary Diagnosis in Surgical Pathology: A Multicenter Blinded Randomized Noninferiority Study of 1992 Cases (Pivotal Study). Am J Surg Pathol 42, 39-52, doi:10.1097/PAS.0000000000000948 (2018).
Kim, J. S., Cho, Y. & Lim, T. H. Prediction of the Location of the Glottis in Laryngeal Images by Using a Novel Deep-Learning Algorithm. Ieee Access 7, 79545-79554, doi:10.1109/Access.2019.2923002 (2019).
Lee, J. Y., Kim, J. S., Kim, T. Y. & Kim, Y. S. Detection and classification of intracranial haemorrhage on CT images using a novel deep-learning algorithm. Sci Rep 10, 20546, doi:10.1038/s41598-020-77441-z (2020).
Kim, M., Kim, J. S., Lee, C. & Kang, B. K. Detection of pneumoperitoneum in the abdominal radiograph images using artificial neural networks. Eur J Radiol Open 8, 100316, doi:10.1016/j.ejro.2020.100316 (2021).
Couture, H. D. et al. Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype. NPJ Breast Cancer 4, 30, doi:10.1038/s41523-018-0079-1 (2018).
Jaber, M. I. et al. A deep learning image-based intrinsic molecular subtype classifier of breast tumors reveals tumor heterogeneity that may affect survival. Breast Cancer Res 22, 12, doi:10.1186/s13058-020-1248-3 (2020).
Liu, T. et al. A Hybrid Deep Learning Model for Predicting Molecular Subtypes of Human Breast Cancer Using Multimodal Data. Irbm 43, 62-74, doi:10.1016/j.irbm.2020.12.002 (2022).
Phan, N. N., Huang, C. C., Tseng, L. M. & Chuang, E. Y. Predicting Breast Cancer Gene Expression Signature by Applying Deep Convolutional Neural Networks From Unannotated Pathological Images. Front Oncol 11, 769447, doi:10.3389/fonc.2021.769447 (2021).
Grossman, R. L. et al. Toward a Shared Vision for Cancer Genomic Data. N Engl J Med 375, 1109-1112, doi:10.1056/NEJMp1607591 (2016).
Cancer Genome Atlas, N. Comprehensive molecular portraits of human breast tumours. Nature 490, 61-70, doi:10.1038/nature11412 (2012).
Allred, D. C., Harvey, J. M., Berardo, M. & Clark, G. M. Prognostic and predictive factors in breast cancer by immunohistochemical analysis. Mod Pathol 11, 155-168 (1998).
Puttagunta, M. & Ravi, S. Medical image analysis based on deep learning approach. Multimed Tools Appl 80, 24365-24398, doi:10.1007/s11042-021-10707-4 (2021).
Riasatian, A. et al. Fine-Tuning and training of densenet for histopathology image representation using TCGA diagnostic slides. Med Image Anal 70, 102032, doi:10.1016/j.media.2021.102032 (2021).
Catak, F. O., Ahmed, J., Sahinbas, K. & Khand, Z. H. Data augmentation based malware detection using convolutional neural networks. PeerJ Comput Sci 7, e346, doi:10.7717/peerj-cs.346 (2021).
Saleh, A. M. & Hamoud, T. Analysis and best parameters selection for person recognition based on gait model using CNN algorithm and image augmentation. J Big Data 8, 1, doi:10.1186/s40537-020-00387-6 (2021).

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Classification of molecular subtypes of breast cancer in whole-slide histopathological images using a deep learning algorithm

Status:

Version 1

Abstract

Figures

INTRODUCTION

RESULT

Classification of molecular subtypes using ANN models

Prediction accuracy of the 120 × 80 pixel model

DISCUSSION

MATERIALS and METHODS

Case selection

Image preprocessing

Training process

Statistical analysis

Declarations

References

Additional Declarations

Status:

Version 1