Application of artificial intelligence in the field of breast pathology diagnosis: narrative review

Introduction

Breast cancer (BC) is a heterogeneous disease with a variable prognosis. It comprises carcinoma in situ (CIS) and invasive disease. The latter can be of no special type (NST) with its wide spectrum of histologic patterns, or it can be classified as one of the special types of BC.

The routine assessment of BC does not employ only histologic typing but also includes biomarker (BM) evaluation by immunohistochemistry (IHC), followed by in situ hybridization (ISH) in a subset of cases. Histologic grading and staging are other important parameters in any complete pathologic report that impact the prognosis and the therapeutic plan.

Digital pathology is a dynamic, rapidly evolving technology that depends on digitalized images rather than the traditional microscopic slides. It emerged as early as the late 1960s as telepathology using digital scanners first invented in 1957. Artificial intelligence (AI) is defined as the ability of a digital computer or computer-controlled robot to perform intelligent tasks commonly performed by the human brain, such as reasoning, finding meaning, generalizing, and learning from past experiences. With the immense advances in translational medicine over decades, AI now shows good to superior concordance rate with conventional slide analysis in various studies (1-7). However, laboratories that implement integrated digital pathology are still limited up to this point. This falls behind other medical disciplines such as radiology, mainly due to the technical limitations related to the bigger size and more detailed pathology images. Manual annotation of digital slides is another hurdle that can be essential sometimes, and it is a time-consuming step.

AI, especially deep learning (DL) has promising results so far. More studies are, however, needed before it is validated for patient care in the clinical setting.

This review is intended to lead to an up-to-date informed use of AI/ML models on breast pathology in both clinical and research settings; and to assist future researchers in the development of new technologies. The author presents this article in accordance with the Narrative Review reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-24-95/rc).

Methods

Information was collected from PubMed and Google Scholar using combined search words like “breast pathology”, “breast biomarkers”, “lymph node metastasis” with “artificial intelligence”, “machine learning”, “deep learning”, and “digital pathology”. Table 1 summarizes the search method.

Background: where are we?

Up to March 2023, more than 520 Food and Drug Administration (FDA) approved AI medical devices are present in the market, 75% of which are related to medical imaging (8).

FDA, Health Canada, UK Medicines and Healthcare Products Regulatory Agency (MHRA) jointly identified 10 guiding principles that can acquaint the development of good machine learning (ML) practice (GMLP) (9). This helps promote safe, effective, high-standard medical devices using AI and ML.

AI is a machine that mimics and overdoes the cognitive capabilities of the human brain. ML is a subfield of AI, where statistical methods learn from data. It can recognize patterns in data with less human instructions. DL is a more developed, complex, multi-layered non-linear end-to-end module with a higher and more abstract representation of ML. It showed successful results in many medical fields such as radiology, diagnosing skin lesions in dermatology, diabetic retinopathy in ophthalmology, and even assessing the detection of microorganisms in medical microbiology (10-12).

To train the machine, a training dataset is provided called input. Examples of correct output or labeling might or might not be provided. ML algorithms can be grouped into four major classes: supervised learning, weakly supervised learning, unsupervised learning, and reinforcement learning. The process is called supervised learning in case the labels were presented. Alternatively, if the training data is provided with a weaker form of annotation and the algorithm is expected to predict the output the method is called weakly supervised learning. This makes the annotation easier and faster, but it has a higher chance of error. Multiple instance learning (MIL) is a subtype of weakly supervised learning. Training data is arranged in bags, each containing a set of instances and labeled, for example, as positive or negative. A weakly supervised approach showed great results in huge datasets (13), but the main problem with this approach is the need for huge datasets. In general, setting more inputs should provide more accurate outcomes (14). On the other hand, in unsupervised learning the algorithm is supposed to analyze and cluster unlabeled datasets. This form of ML can potentially figure out patterns and associations that were not known before and detect novel discoveries. Lastly, in reinforcement learning ML is feedback-based (15,16). Supervised ML other than DL such as decision tree, random forest, support vector machine (SVM), and weakly supervised algorithms like MIL were also used in breast pathology image analysis efficiently with easy training.

In DL, the first layer takes the input data and the last layer produces the output. In between, several hidden layers of neural networks perform complex operations. Non-DL machines require structural data and use traditional algorithms like linear regression. On the contrary, DL takes both structural and unstructured data and uses neural networks to filter and make brain-like decisions. DL is more effective if abundant labeled training data is obtainable, and the trained DL network can act like a representation method.

The algorithm of ML can detect and classify tumors, recognize the phenotype akin to that of genomics, and predict therapy response. Whole slide images (WSIs) are considered the black boxes of AI, as they have high resolution and provide enormous input data. WSIs became more popular on a large scale in clinical practice after the FDA approved the WSI scanners in 2017 (17). Before an algorithm can use the data to learn a model, representation of the potentially relevant data and discard of the irrelevant distractors is needed. This step was used to be done manually. Recently, representation learning can be done with a convolutional neural network (CNN or convent) and the machine may automatically detect the relevant dataset. CNN is a subset of ML, and precisely DL. It is one of the various types of artificial neural networks used for different applications and data types. CNN-based algorithm detects 51 features associated with BC, including cytologic and architectural characteristics of tumor cells, and other attributes such as inflammation and microcalcification (18). Unsupervised domain adapts self-supervised learning and the machine learns morphological, geometric, and contextual contents of images using unlabeled data (19). CNN is the most utilized type of DL in breast pathology.

Generative adversarial network (GAN) is another special type of DL in which the algorithm has 2 domains, the first is responsible for generating data and the other one determines if the output looks real.

Applications of AI in digital breast pathology

• Quantification of BM and Ki67.
• Mitoses detection.
• Lymph node metastasis recognition.
• Computer-aided diagnosis and classification of in situ lesions and invasive BC.
• Prognostication.
• Assessment of the surrounding stroma and tumor-infiltrated lymphocytes (TILs).
• Prediction of molecular expression and treatment response outcome.
• Grading of BC.
• Frozen section and rapid diagnosis.

Quantification of BM and Ki67

The IHC-based assessment of breast BM including estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), along with the evaluation of the Ki67 proliferation index is very important as a prognostic and predictive indicator. ER determines the inclusion of endocrine therapy and the PR reflects its good response (20,21). The cutoff point of 1% nuclear positivity according to the guidelines of the College of American Pathologists (CAP) is routinely assessed manually. The amplification of HER2 oncogene is associated with poor prognosis and it is an indication to start anti-HER2 targeted therapy. Reproducibility of BM assessment in general and HER2 in particular is a constrain.

The BM assessment by IHC is frequently used to approximate the genetic profile. The gene expression profile is the genetic signature of BC. It adds prognostic information, and it can classify BC into high- and low-risk subtypes (22,23). Molecular assays are available in the market under various trade names like BluePrint, MammaPrint, Prosigna, Oncotype DX. They help to define the subtype and estimate the recurrence risk. However, they are not equally obtainable for everyone and often hard to afford. Improving the BM assessment by automation can significantly enhance the patient’s care. Image J is a freely, open java-based IA, developed by the National Institute of Health (Bethesda, MD, USA) and it can be used to quantify the IHC markers in pathology including the breast BM (24). Immunoratio is another AI tool devoted to breast BM assessment (for ER, PR, and Ki67), that is available as a web application, or as a plugin for image J, and it shows an excellent concordance for Ki67 conventional manual assessment in BC (25,26). Likewise, ImmunoMembrane is used for HER2 assessment in BC (27). AstraZeneca is incorporating the data analysis company Definiens “Tissue Phenomics software” for predicting new BM discoveries (28-30). DL algorithms have also been applied to automated BM assessment in BC images, including automated ER (31) and HER2 (32). Visiopharm is an AI that has many utilities in breast pathology including BM assessment, TIL quantification, and detection of nodal metastasis (33-35). Hence, it may efficiently facilitate the workflow. Aperio Digital Pathology and TissueGnostics Analysis are other available software for breast BM assessment (36). Most automation of breast pathology focused on BM analysis. Automated BM evaluation in BC has the power of high reproducibility and accordingly, it can act as an accountable reflection of gene expression assays. Moreover, there is a potential decline of the human epidermal growth factor receptor 2-in situ hybridization (HER2-ISH) requisite as the equivocal cases decrease with the automation (37,38). Overall, digitalization provides an efficient, economical, faster, and more reproducible way for BM evaluation through computerized image analysis.

Likewise, the proliferation index Ki67 is usually assessed by manual counting of the positive proportion of at least 200 tumor cells in a hotspot. The current guidelines recommend that at least three representative high-power fields (HPFs) must be selected. Poor reproducibility results from discrepancy in field selection between hotspot versus average scoring, tumor cell counting, and possible undercounting of the faint positive nuclei (39). Therefore, despite the clinically validated value of Ki67 as a prognostic and predictive factor in BC, its manual assessment is considered far from standardization (39-41).

Studies have shown that digital image analysis might help to reduce the variability in Ki67 scoring. A recent study of automated scoring of Ki-67 staining employing Mindpeak software demonstrated remarkable potential as the issues of lack of consistency, reproducibility, and accuracy can be eliminated (42).

Automated scoring has the benefits of being faster and more objective, suggesting that it could potentially improve diagnostic reproducibility, minimizing the inter-observer variability, and, hence, facilitate clinical decision-making.

Mitoses detection

The mitotic cell count is an important parameter for grading BC, revelation of the prognosis, and prediction of both the tumor aggressiveness and therapy response. Typically, pathologists examine histopathology images manually under high-resolution microscopes to count mitoses in a subjective and time-consuming process. Initial automation of mitoses detection based on AI showed low accuracy and high computational cost. Mahmood et al. developed a multistage mitotic-cell-detection method based on a faster region convolutional neural network (Faster R-CNN) and deep CNNs. They obtained promising and generalized results utilizing two open datasets [the International Conference on Pattern Recognition (ICPR) 2012 and ICPR 2014 (MITOS-ATYPIA-14)] (43). Likewise, Shihabuddin and K showed superior performance of multi-CNN consisting of 3 pre-trained CNNs with linear SVM in comparison to multi-CNN combinations with other classifiers such as Adaboost and random forest (44). A recent pilot study by van Bergeijk using ML showed that mitotic count on WSI correlates well with AI-assisted mitotic count and it can be safely used in clinical settings (45).

Lymph node metastasis recognition

CNN can act as a screening tool on sentinel lymph node WSI stained by pan-cytokeratin IHC with a high accuracy rate and 100% sensitivity on negative cases (28,37,46-49). In a review of 23 experimental studies published by Caldonazzi et al., 55% of used algorithms; including all 4 weekly supervised studies achieved an efficiency performance of >95% across all parameters including precision, accuracy, sensitivity, specificity, and area under the curve of the receiver operating curves (47).

Computer-aided diagnosis and classification of in situ lesions and invasive BC

In general, utilizing WSI demonstrated an overall concordance rate of 97.1% in comparison with the conventional slides for breast needle biopsies (50).

In an experiment done by Gecer et al. applying CNN on WSI to categorize breast images into 5 diagnostic classes, DL performed better than competing methods that used hand-crafted features and statistical classifiers (51). A DL algorithm developed by Mehta et al. showed matching to the classification accuracy of 87 U.S. pathologists for a challenging test set (52).

DL algorithms can help in classifying an intraductal lesion to usual or atypical duct hyperplasia (UDH & ADH; respectively), or ductal carcinoma in situ (DCIS) (53), and accordingly select the patients who need to proceed to surgical excision. DL can also stratify cases of DCIS according to nuclear grade and probability of progressing to invasive BC.

More studies, however, focus on invasive BC detection and classification than the in-situ lesions (54,55).

In general pathology, Kappa values of several investigates showed greater or even metrics of AI agreement compared to pathologists (56-59). The external validation of AI algorithm application developed by Ibex Galen for diagnosing BC in biopsies demonstrated both clinical utility and accuracy in real-world clinical use for various types of invasive BC, and both low- and high-grade DCIS (58).

Computer-aided prognostication

The prognostic models on BC are usually based on clinical, histologic, and molecular characteristics (60,61), and AI might improve the reproducibility of all. With utilizing ML, new stromal features not previously known to have an association with prognosis in BC were defined and can be employed to predict the patients’ outcome and response to therapy (62,63).

Edge-based feature ML was tested to predict BC metastasis and showed superior results in comparison with other forms of ML (64).

Additionally, the hypoxia-related dysregulation of mRNA and microRNAs was identified to predict the prognosis of BC through co-expression networks and clustering (65).

A study of the clinical, pathological, and biological variables of 160 cases of metaplastic BC employing five models (namely, decision tree with bagging, logistic regression, multilayer perceptron, naive Bayes, and random forest algorithms) revealed that the random forest model exhibited the highest performance in predicting BC specific survival, emphasizing the potential of ML algorithms in predicting prognosis for complex and heterogeneous BC subtypes (66). Various graph-neural networks (GNNs) models were trained in Kim’s study using cancer patients’ genomic and clinical data with promising outcomes (67).

Assessment of the surrounding stroma and TIL

The tumor-associated stroma tissue is understudied in conventional breast pathology. Both digital tumor-associated stroma score and TIL score were found to carry strong prognostic value for disease-specific survival in triple-negative BC using CNN (62). Likewise, AI has been developed to quantify and spatially analyze immune cells, proportionate stroma, and detect tumor budding.

It showed that collagenous stroma on WSI was best associated with lower rates of pathologic complete response (pCR), while combined high proportionated stroma (myxoid, collagenous, and immune) most optimally predicted worse clinical survival outcomes (63). Studying tumor-associated fibroblasts by ML was also done to construct a prognostic model for directing immunotherapy in patients with BC (24).

TILs are a reliable and reproducible marker of tumor immunogenicity in BC. Higher levels of TILs are associated with better prognosis in the early stage, especially non-luminal BC, and a higher probability of achieving complete response in the neoadjuvant setting. Analysis of TILs in residual disease specimens after neoadjuvant therapy has also been shown to have prognostic value. Despite its growing recognition as an immunologic BM in BC, TIL-quantification is a lengthy process that requires accurate single-cell level segmentation on pathological images. This limits both its application and accuracy without the help of AI. Currently, the evaluation of TIL is a semi-quantitative, manual, and lengthy process. By utilizing AI/ML, the standardization and accuracy would improve. However, it is still experimental till now and not sufficiently validated for standard practice. In Thagaard et al.’s review, many pitfalls were detected as sources of discrepancy between visual and computational assessments (68). These include incorrect foci of evaluation, heterogeneity in distribution as TILs are usually more at the leading edge of a tumor compared to the central tumor area, and tumor-stroma outlining becomes too precise beyond human eye capabilities. Technical factors related to the slides also are common pitfall sources, including those derived from tissue processing, microtomy, staining, and mounting (zonal fixation, blade lines, tissue disruption, microchatters, air bubbles, floaters). Out-of-focus areas, pen markings, tissue folds, blurring, air bubbles, thick sections, and crush and cautery artifacts can each confuse the machine and consequently lead to inaccurate quantification. Further, challenges related to image analysis while adhering to a clinical guideline and training data to create robust, validated, and generalizable algorithms are all influencers of the clinical application in the routine workflow. Promising results related to TIL assessment were demonstrated in a recent validated study (58).

Prediction of molecular expression and treatment response outcome

DL can predict the subtypes from genetic information more accurately than conventional statistical methods. A DL model called point-wise linear (PWL) generates a custom-made logistic regression for patients and reveals the relationships between the 50-gene signature that classifies BC into 5 molecular intrinsic subtypes and the copy numbers of BC. The PWL model utilized genes relevant to the cell cycle-related pathways and showed preliminary successes in BC subtype analysis, demonstrating the potential to clarify the mechanisms underlying BC and improve overall clinical outcomes (69).

CNN was used to evaluate gene expression profiles collected from The Cancer Genome Atlas (TCGA) database (70). Likewise, DL was used to explore the subtype-specific expression pattern, whereby 6 new subtypes of triple-negative BC were discovered (71).

A multi-omic ML predictor of BC therapy response was developed using a combination of pre-therapy features, including tumor mutational and copy number landscapes, tumor proliferation, immune infiltration, and T cell dysfunction and exclusion (72). Also, ML models using image-based features derived from tumor immune micro-environment, BMs, and clinical features accurately predicted the response to neoadjuvant chemotherapy in non-luminal BC patients and outperformed the results learned by pathologists’ manually assessed features (73).

Moreover, gene expression and copy number alternations were analyzed via a modified DL algorithm, and 6 subgroups in HER2-positive groups were found (74).

Grading

DL can detect both tubular formation (74) and nuclear pleomorphism (75) and count mitotic figures (76-81). Hence, it can provide accurate histologic tumor grading. A great advantage of machine-assisted grading is reducing subjectivity. Besides, it can be time-saving.

Frozen section and rapid diagnosis

Although to start with, the utility of digital pathology in intra-operative consultation is not common. The experience of the University Health Network using robotic microscopy and virtual slide telepathology in frozen section diagnosis for all systems showed a 98% diagnostic accuracy rate, and 7.7% overall deferral rate, with mid-case technical failure causing delay has occurred in only 3 cases (0.3%) (6). In another study comparing WSI and conventional microscopy to identify tumors in en face frozen sections for skin cases the agreement between the 2 modalities was 100% but the WSI assessment using Aperio analysis was time-consuming (82).

Dynamic full-field optical coherence tomography (D-FF-OCT) microscopy provides fast, high-resolution images of excised tissues with a contrast comparable to hematoxylin and eosin (H&E) stain histology but without any tissue preparation and alteration. Combined with ML, a high diagnostic accuracy rate was obtained in a pilot study on 51 cases of BC (83).

DL algorithm was also developed to assess sentinel lymph node metastasis, a common indication of frozen section in the setting of BC. Micrometastasis was the only challenge that was found to lower the accuracy rate in Kim et al. study (48); while micro-metastatic tumors, neoadjuvant systemic therapy, invasive lobular carcinoma, and histologic grade 3 were associated with lower diagnostic accuracy in another study (49). Overall, published data suggests that the application of AI in detecting nodal metastases is promising and can be competently employed in daily pathology practice (47).

Advantages/benefits of AI

In the era of personalized therapy that usually requires knowledge of the genetic makeup of the tumor, we face a prompt and crucial necessity to have more precise, accurate, objective and swift assessment and typing of BC. AI makes this attainable. Computerized image analysis enhances objectivity and facilitates obtaining consultation, consensus, and expert opinion. Automated BM assessment is faster and more reproducible which mitigates the workflow. Digital images are more durable and easier to save for longer periods than traditional slides which faint over time. A cardinal benefit of AI and ML is the detection of patterns and potential associations that predict the prognosis or genetic makeup of the tumor that are not visible to the naked eye and the conventional microscope. Overall, ML showed high concordance in BC diagnosis, especially for invasive cases (18). Hence, increasing accuracy rate, evading human error, and alleviating the workflow to keep with the demands of the upcoming era of personalized therapy in a cost and money-effective manner are all goals of utilizing AI in breast pathology service.

Disadvantages/limitations of AI

AI is relatively slowly evolving in the medical field in general probably because of the high cost and the dependence on robust information technology (IT) support systems. In breast pathology, the encouraging results were mainly studied on breast WSI of resection specimens rather than on core biopsies; the regular sample used for diagnosis. Using WSI has its drawbacks, such as slide scan failure, prolonged time for pathologists to review cases, and a need for higher image resolution (84).

In AI, technical issues are there, as the learning strategy in DL might not be immediately optimal for mitigating data, and the data scale of many studies is below the common standards of ML (28). The standard DL architectures are usually expensive and require huge memory especially while training the machine. Supervised learning does better than weakly supervised and unsupervised learning. The need for annotation in supervised learning is a hitch that can be costly and time and effort-consuming. Besides, supervised learning deprives the machine of the big advantage of novel discoveries that can emerge with unsupervised learning on WSI. Domain shift and system reliability are major concerns if the encountered data distribution varies from the training dataset. Successful algorithms for tumor classification and prognosis might do poorly when applied by an external laboratory, which is a major limitation. Multi-center analysis is crucial to overcome the possibility of domain shift. The need for a massive number of cases with manual annotations is a major strain on the validation process. GAN can generate realistic data augmenting the dataset and enhancing the machine performance with limited training. AI-based cell phenotyping is highly dependent on extracted information from the H&E slides. Therefore, AI analysis is subjected to various pre-analytical and analytic factors including section thickness, processing, staining quality, and the scanner used. Pro-processing conditional GAN can translate images from one form to another resulting in high-quality translated images, and it can be utilized to normalize the colors (85). Janowczyk and his group have developed an open-source quality control tool (HistoQC) for digital pathology slides to address the issue of staining quality control (86). Additionally, ML is very sensitive. It can detect everything on the slide including fingerprints, scratches, ink dots, and dust marks. False positive output might be produced by histologic lesions such as necrosis, fibroadenomatoid changes, and hyperplasia (18).

Internal and external validation of AI models with complete transparency is of paramount importance and it is considered by some researchers as a false hope (87).

Ethical AI is an important concept to be adhered to all the time. This includes protecting autonomy, ensuring transparency and patient safety, accountability, inclusiveness, and governance (88). Besides, pathologists should be aware of potential bias and system pitfalls as limitations of using AI in practice.

Conclusions

Up to date, only a few laboratories have adapted and integrated a digital workflow which hinders the maturity of AI applications. Despite the documentation that automated image analysis improved the reproducibility of BM assessment and DL algorithms were effective for computer-assisted prognosis based on molecular subtyping and grading, exciting results will not transpire early. Currently, the landscape of the usage of AI varies from promising success such as quantification of BM, Ki67 and mitotic counts; evaluation of lymph node metastasis in routine and frozen sections; assessment of grading and the surrounding stroma; and predicting therapy response and prognostication, to yet immature potential applications such as TIL quantification and diagnosis of breast lesions in frozen sections. Despite all the challenges, integration of AI in the clinical practice of breast pathology is eventually needed to move forward as it aligns with personalized and precision medicine.

Acknowledgments

Funding: None.

Footnote

Reporting Checklist: The author has completed the Narrative Review reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-24-95/rc

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-24-95/prf

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-24-95/coif). The author has no conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ho J, Parwani AV, Jukic DM, et al. Use of whole slide imaging in surgical pathology quality assurance: design and pilot validation studies. Hum Pathol 2006;37:322-31. [Crossref] [PubMed]
2. Helin H, Lundin M, Lundin J, et al. Web-based virtual microscopy in teaching and standardizing Gleason grading. Hum Pathol 2005;36:381-6. [Crossref] [PubMed]
3. Fine JL, Grzybicki DM, Silowash R, et al. Evaluation of whole slide image immunohistochemistry interpretation in challenging prostate needle biopsies. Hum Pathol 2008;39:564-72. [Crossref] [PubMed]
4. López AM, Graham AR, Barker GP, et al. Virtual slide telepathology enables an innovative telehealth rapid breast care clinic. Hum Pathol 2009;40:1082-91. [Crossref] [PubMed]
5. Velez N, Jukic D, Ho J. Evaluation of 2 whole-slide imaging applications in dermatopathology. Hum Pathol 2008;39:1341-9. [Crossref] [PubMed]
6. Evans AJ, Chetty R, Clarke BA, et al. Primary frozen section diagnosis by robotic microscopy and virtual slide telepathology: the University Health Network experience. Hum Pathol 2009;40:1070-81. [Crossref] [PubMed]
7. Wienert S, Beil M, Saeger K, et al. Integration and acceleration of virtual microscopy as the key to successful implementation into the routine diagnostic process. Diagn Pathol 2009;4:3. [Crossref] [PubMed]
8. Sinha R. Over 500 Healthcare AI Algorithms Now Approved by FDA. March 2, 2023. Accessed in 01/01/2024. Available online: https://collaborativeimaging.com/2023/03/02/over-500-healthcare-ai-algorithms-now-approved-by-fda/#:~:text=The%20FDA%20recently%20announced%20that,first%20AI%20algorithm%20in%201995
9. Good Machine Learning Practice for Medical Device Development: Guiding Principles. 2021. Accessed in 24/10/2023. Available online: https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles
10. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017;542:115-8. Erratum in: Nature 2017;546:686. [Crossref] [PubMed]
11. Gulshan V, Peng L, Coram M, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA 2016;316:2402-10. [Crossref] [PubMed]
12. Marletta S, L'Imperio V, Eccher A, et al. Artificial intelligence-based tools applied to pathological diagnosis of microbiological diseases. Pathol Res Pract 2023;243:154362. [Crossref] [PubMed]
13. Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med 2019;25:1301-9. [Crossref] [PubMed]
14. Koteluk O, Wartecki A, Mazurek S, et al. How Do Machines Learn? Artificial Intelligence as a New Era in Medicine. J Pers Med 2021;11:32. [Crossref] [PubMed]
15. Reinders C, Ackermann H, Yang MY, et al. Chapter 4 - Learning Convolutional Neural Networks for Object Detection with Very Little Training Data. In: Yang MY, Rosenhahn B, Murino V. editors. Multimodal Scene Understanding. Academic Press; 2019:65-100.
16. Jonsson A. Deep Reinforcement Learning in Medicine. Kidney Dis (Basel) 2019;5:18-22. [Crossref] [PubMed]
17. FDA. FDA allows marketing of first whole slide imaging system for digital pathology. 2017. Accessed in 30/12/2023. Available online: https://www.fda.gov/news-events/press-announcements/fda-allows-marketing-first-whole-slide-imaging-system-digital-pathology
18. Badve SS. Artificial intelligence in breast pathology - dawn of a new era. NPJ Breast Cancer 2023;9:5. [Crossref] [PubMed]
19. Jose L, Liu S, Russo C, et al. Generative Adversarial Networks in Digital Pathology and Histopathological Image Processing: A Review. J Pathol Inform 2021;12:43. [Crossref] [PubMed]
20. Barnes DM, Hanby AM. Oestrogen and progesterone receptors in breast cancer: past, present and future. Histopathology 2001;38:271-4. [Crossref] [PubMed]
21. Harris LN, Ismaila N, McShane LM, et al. Use of Biomarkers to Guide Decisions on Adjuvant Systemic Therapy for Women With Early-Stage Invasive Breast Cancer: American Society of Clinical Oncology Clinical Practice Guideline. J Clin Oncol 2016;34:1134-50. [Crossref] [PubMed]
22. Sørlie T, Perou CM, Tibshirani R, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 2001;98:10869-74. [Crossref] [PubMed]
23. Perou CM, Sørlie T, Eisen MB, et al. Molecular portraits of human breast tumours. Nature 2000;406:747-52. [Crossref] [PubMed]
24. Robertson S, Azizpour H, Smith K, et al. Digital image analysis in breast pathology-from image processing techniques to artificial intelligence. Transl Res 2018;194:19-35. [Crossref] [PubMed]
25. Tuominen VJ, Ruotoistenmäki S, Viitanen A, et al. ImmunoRatio: a publicly available web application for quantitative image analysis of estrogen receptor (ER), progesterone receptor (PR), and Ki-67. Breast Cancer Res 2010;12:R56. [Crossref] [PubMed]
26. Yeo MK, Kim HE, Kim SH, et al. Clinical usefulness of the free web-based image analysis application ImmunoRatio for assessment of Ki-67 labelling index in breast cancer. J Clin Pathol 2017;70:715-9. [Crossref] [PubMed]
27. Tuominen VJ, Tolonen TT, Isola J. ImmunoMembrane: a publicly available web application for digital image analysis of HER2 immunohistochemistry. Histopathology 2012;60:758-67. [Crossref] [PubMed]
28. Stålhammar G, Fuentes Martinez N, Lippert M, et al. Digital image analysis outperforms manual biomarker assessment in breast cancer. Mod Pathol 2016;29:318-29. [Crossref] [PubMed]
29. Lykkegaard Andersen N, Brügmann A, Lelkaitis G, et al. Virtual Double Staining: A Digital Approach to Immunohistochemical Quantification of Estrogen Receptor Protein in Breast Carcinoma Specimens. Appl Immunohistochem Mol Morphol 2018;26:620-6. [Crossref] [PubMed]
30. Røge R, Riber-Hansen R, Nielsen S, et al. Proliferation assessment in breast carcinomas using digital image analysis based on virtual Ki67/cytokeratin double staining. Breast Cancer Res Treat 2016;158:11-9. [Crossref] [PubMed]
31. Mungle T, Tewary S, Das DK, et al. MRF-ANN: a machine learning approach for automated ER scoring of breast cancer immunohistochemical images. J Microsc 2017;267:117-29. [Crossref] [PubMed]
32. Vandenberghe ME, Scott ML, Scorer PW, et al. Relevance of deep learning to facilitate the diagnosis of HER2 status in breast cancer. Sci Rep 2017;7:45938. [Crossref] [PubMed]
33. Thagaard J, Stovgaard ES, Vognsen LG, et al. Automated Quantification of sTIL Density with H&E-Based Digital Image Analysis Has Prognostic Potential in Triple-Negative Breast Cancers. Cancers (Basel) 2021;13:3050. [Crossref] [PubMed]
34. Hartage R, Li AC, Hammond S, et al. A Validation Study of Human Epidermal Growth Factor Receptor 2 Immunohistochemistry Digital Imaging Analysis and its Correlation with Human Epidermal Growth Factor Receptor 2 Fluorescence In situ Hybridization Results in Breast Carcinoma. J Pathol Inform 2020;11:2. [Crossref] [PubMed]
35. Challa B, Tahir M, Hu Y, et al. Artificial Intelligence-Aided Diagnosis of Breast Cancer Lymph Node Metastasis on Histologic Slides in a Digital Workflow. Mod Pathol 2023;36:100216. [Crossref] [PubMed]
36. Ahern TP, Beck AH, Rosner BA, et al. Continuous measurement of breast tumour hormone receptor expression: a comparison of two computational pathology platforms. J Clin Pathol 2017;70:428-34. [Crossref] [PubMed]
37. Holten-Rossing H, Møller Talman ML, Kristensson M, et al. Optimizing HER2 assessment in breast cancer: application of automated image analysis. Breast Cancer Res Treat 2015;152:367-75. [Crossref] [PubMed]
38. Brügmann A, Eld M, Lelkaitis G, et al. Digital image analysis of membrane connectivity is a robust measure of HER2 immunostains. Breast Cancer Res Treat 2012;132:41-9. [Crossref] [PubMed]
39. Al Nemer A. The Performance of Ki-67 Labeling Index in Different Specimen Categories of Invasive Ductal Carcinoma of the Breast Using 2 Scoring Methods. Appl Immunohistochem Mol Morphol 2017;25:86-90. [Crossref] [PubMed]
40. Nielsen TO, Leung SCY, Rimm DL, et al. Assessment of Ki67 in Breast Cancer: Updated Recommendations From the International Ki67 in Breast Cancer Working Group. J Natl Cancer Inst 2021;113:808-19. [Crossref] [PubMed]
41. Curigliano G, Burstein HJ, Winer EP, et al. De-escalating and escalating treatments for early-stage breast cancer: the St. Gallen International Expert Consensus Conference on the Primary Therapy of Early Breast Cancer 2017. Ann Oncol 2017;28:1700-12. Erratum in: Ann Oncol 2018;29:2153; Ann Oncol 2019;30:1181. [Crossref] [PubMed]
42. Zehra T, Shams M, Ahmad Z, et al. Ki-67 Quantification in Breast Cancer by Digital Imaging AI Software and its Concordance with Manual Method. J Coll Physicians Surg Pak 2023;33:544-7. [Crossref] [PubMed]
43. Mahmood T, Arsalan M, Owais M, et al. Artificial Intelligence-Based Mitosis Detection in Breast Cancer Histopathology Images Using Faster R-CNN and Deep CNNs. J Clin Med 2020;9:749. [Crossref] [PubMed]
44. Shihabuddin AR. K SB. Multi CNN based automatic detection of mitotic nuclei in breast histopathological images. Comput Biol Med 2023;158:106815. [Crossref] [PubMed]
45. van Bergeijk SA, Stathonikos N, Ter Hoeve ND, et al. Deep learning supported mitoses counting on whole slide images: A pilot study for validating breast cancer grading in the clinical workflow. J Pathol Inform 2023;14:100316. [Crossref] [PubMed]
46. Litjens G, Sánchez CI, Timofeeva N, et al. Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Sci Rep 2016;6:26286. [Crossref] [PubMed]
47. Caldonazzi N, Rizzo PC, Eccher A, et al. Value of Artificial Intelligence in Evaluating Lymph Node Metastases. Cancers (Basel) 2023;15:2491. [Crossref] [PubMed]
48. Kim YG, Song IH, Cho SY, et al. Diagnostic Assessment of Deep Learning Algorithms for Frozen Tissue Section Analysis in Women with Breast Cancer. Cancer Res Treat 2023;55:513-22. [Crossref] [PubMed]
49. Kim YG, Song IH, Lee H, et al. Challenge for Diagnostic Assessment of Deep Learning Algorithm for Metastases Classification in Sentinel Lymph Nodes on Frozen Tissue Section Digital Slides in Women with Breast Cancer. Cancer Res Treat 2020;52:1103-11. [Crossref] [PubMed]
50. Campbell WS, Hinrichs SH, Lele SM, et al. Whole slide imaging diagnostic concordance with light microscopy for breast needle biopsies. Hum Pathol 2014;45:1713-21. [Crossref] [PubMed]
51. Gecer B, Aksoy S, Mercan E, et al. Detection and classification of cancer in whole slide breast histopathology images using deep convolutional networks. Pattern Recognit 2018;84:345-56. [Crossref] [PubMed]
52. Mehta S, Lu X, Wu W, et al. End-to-End diagnosis of breast biopsy images with transformers. Med Image Anal 2022;79:102466. [Crossref] [PubMed]
53. Dundar MM, Badve S, Bilgin G, et al. Computerized classification of intraductal breast lesions using histopathological images. IEEE Trans Biomed Eng 2011;58:1977-84. [Crossref] [PubMed]
54. Araújo T, Aresta G, Castro E, et al. Classification of breast cancer histology images using Convolutional Neural Networks. PLoS One 2017;12:e0177544. [Crossref] [PubMed]
55. Cruz-Roa A, Gilmore H, Basavanhally A, et al. Accurate and reproducible invasive breast cancer detection in whole-slide images: A Deep Learning approach for quantifying tumor extent. Sci Rep 2017;7:46450. [Crossref] [PubMed]
56. Bulten W, Pinckaers H, van Boven H, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study. Lancet Oncol 2020;21:233-41. [Crossref] [PubMed]
57. Nagpal K, Foote D, Liu Y, et al. Development and validation of a deep learning algorithm for improving Gleason scoring of prostate cancer. NPJ Digit Med 2019;2:48. [Crossref] [PubMed]
58. Sandbank J, Bataillon G, Nudelman A, et al. Validation and real-world clinical application of an artificial intelligence algorithm for breast cancer detection in biopsies. NPJ Breast Cancer 2022;8:129. [Crossref] [PubMed]
59. Ström P, Kartasalo K, Olsson H, et al. Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study. Lancet Oncol 2020;21:222-32. [Crossref] [PubMed]
60. Madabhushi A, Agner S, Basavanhally A, et al. Computer-aided prognosis: predicting patient and disease outcome via quantitative fusion of multi-scale, multi-modal data. Comput Med Imaging Graph 2011;35:506-14. [Crossref] [PubMed]
61. Chen JM, Li Y, Xu J, et al. Computer-aided prognosis on breast cancer with hematoxylin and eosin histopathology images: A review. Tumour Biol 2017;39:1010428317694550. [Crossref] [PubMed]
62. Albusayli R, Graham JD, Pathmanathan N, et al. Artificial intelligence-based digital scores of stromal tumour-infiltrating lymphocytes and tumour-associated stroma predict disease-specific survival in triple-negative breast cancer. J Pathol 2023;260:32-42. [Crossref] [PubMed]
63. Hacking SM, Karam J, Singh K, et al. Whole slide image features predict pathologic complete response and poor clinical outcomes in triple-negative breast cancer. Pathol Res Pract 2023;246:154476. [Crossref] [PubMed]
64. Adnan N, Lei C, Ruan J. Robust edge-based biomarker discovery improves prediction of breast cancer metastasis. BMC Bioinformatics 2020;21:359. [Crossref] [PubMed]
65. Gong PJ, Shao YC, Huang SR, et al. Hypoxia-Associated Prognostic Markers and Competing Endogenous RNA Co-Expression Networks in Breast Cancer. Front Oncol 2020;10:579868. [Crossref] [PubMed]
66. Feng Y, McGuire N, Walton A, et al. Predicting breast cancer-specific survival in metaplastic breast cancer patients using machine learning algorithms. J Pathol Inform 2023;14:100329. [Crossref] [PubMed]
67. Kim SY. GNN-surv: Discrete-Time Survival Prediction Using Graph Neural Networks. Bioengineering (Basel) 2023;10:1046. [Crossref] [PubMed]
68. Thagaard J, Broeckx G, Page DB, et al. Pitfalls in machine learning-based assessment of tumor-infiltrating lymphocytes in breast cancer: A report of the International Immuno-Oncology Biomarker Working Group on Breast Cancer. J Pathol 2023;260:498-513. [Crossref] [PubMed]
69. Shibahara T, Wada C, Yamashita Y, et al. Deep learning generates custom-made logistic regression models for explaining how breast cancer subtypes are classified. PLoS One 2023;18:e0286072. [Crossref] [PubMed]
70. Mostavi M, Chiu YC, Huang Y, et al. Convolutional neural network models for cancer type prediction based on gene expression. BMC Med Genomics 2020;13:44. [Crossref] [PubMed]
71. Liu J, Su R, Zhang J, et al. Classification and gene selection of triple-negative breast cancer subtype embedding gene connectivity matrix in deep neural network. Brief Bioinform 2021;22:bbaa395. [Crossref] [PubMed]
72. Sammut SJ, Crispin-Ortuzar M, Chin SF, et al. Multi-omic machine learning predictor of breast cancer therapy response. Nature 2022;601:623-9. [Crossref] [PubMed]
73. Li Z. Artificial intelligence predicts breast cancer neoadjuvant chemotherapy response from multi-source pathologic images. Journal of Pathology Informatics 2022;13:100040. [Crossref]
74. Mohaiminul Islam M, Huang S, Ajwad R, et al. An integrative deep learning framework for classifying molecular subtypes of breast cancer. Comput Struct Biotechnol J 2020;18:2185-99. [Crossref] [PubMed]
75. Dalle JR, Leow WK, Racoceanu D, et al. Automatic breast cancer grading of histopathological images. Annu Int Conf IEEE Eng Med Biol Soc 2008;2008:3052-5. [Crossref] [PubMed]
76. Cireşan DC, Giusti A, Gambardella LM, et al. Mitosis detection in breast cancer histology images with deep neural networks. Med Image Comput Comput Assist Interv 2013;16:411-8.
77. Wang H, Cruz-Roa A, Basavanhally A, et al. Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features. J Med Imaging (Bellingham) 2014;1:034003. [Crossref] [PubMed]
78. Veta M, van Diest PJ, Willems SM, et al. Assessment of algorithms for mitosis detection in breast cancer histopathology images. Med Image Anal 2015;20:237-48. [Crossref] [PubMed]
79. Malon CD, Cosatto E. Classification of mitotic figures with convolutional neural networks and seeded blob features. J Pathol Inform 2013;4:9. [Crossref] [PubMed]
80. Irshad H. Automated mitosis detection in histopathology using morphological and multi-channel statistics features. J Pathol Inform 2013;4:10. [Crossref] [PubMed]
81. Roux L, Racoceanu D, Loménie N, et al. Mitosis detection in breast cancer histological images An ICPR 2012 contest. J Pathol Inform 2013;4:8. [Crossref] [PubMed]
82. Kim YH, Pham KA, Castillo SA, et al. Comparing Whole Slide Imaging and Light Microscopy to Identify Tumor in en face Frozen Sections. Ann Clin Lab Sci 2022;52:986-90. [PubMed]
83. Scholler J, Mandache D, Mathieu MC, et al. Automatic diagnosis and classification of breast surgical samples with dynamic full-field OCT and machine learning. J Med Imaging (Bellingham) 2023;10:034504. [Crossref] [PubMed]
84. Rizzo PC, Girolami I, Marletta S, et al. Technical and Diagnostic Issues in Whole Slide Imaging Published Validation Studies. Front Oncol 2022;12:918580. [Crossref] [PubMed]
85. Khan AM, Rajpoot N, Treanor D, et al. A nonlinear mapping approach to stain normalization in digital histopathology images using image-specific color deconvolution. IEEE Trans Biomed Eng 2014;61:1729-38. [Crossref] [PubMed]
86. Janowczyk A, Zuo R, Gilmore H, et al. HistoQC: An Open-Source Quality Control Tool for Digital Pathology Slides. JCO Clin Cancer Inform 2019;3:1-7. [Crossref] [PubMed]
87. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health 2021;3:e745-50. [Crossref] [PubMed]
88. Chauhan C, Gullapalli RR. Ethics of AI in Pathology: Current Paradigms and Emerging Issues. Am J Pathol 2021;191:1673-83. [Crossref] [PubMed]