Artifical Intelligence in Breast Cancer Pathology

 

Introduction

Breast cancer is the most commonly diagnosed cancer in women, and the biopsy slide is still the final word on it. That last step is also the slowest and the most subjective. Reading histopathology takes years of training, and even experienced pathologists disagree on the harder calls, especially at the fuzzy borders between benign, atypical, and malignant tissue. Caseloads keep rising while the number of specialists does not. That pressure is what has moved artificial intelligence from a research idea into a tool that increasingly sits beside the pathologist at the microscope. Looking across the recent literature, a few clear shifts stand out.


From glass slides to digital ones

                                      Figure: WSI image of breast cancer with 10x, 20x ,and 40x magnification.

None of this happens without digitization. Whole slide imaging, which turns a glass slide into a high-resolution digital file, is the foundation on which everything else is built, and it has quietly become accepted practice rather than an experiment. Regulators, including the FDA, have cleared digital pathology for clinical use, and large studies show that reading a slide on a screen is diagnostically equivalent to reading it under a microscope. Only once a slide is a file can an algorithm work on it, so this quiet shift is what made everything after it possible.

From detection to decision support

The most mature use of AI is in diagnosis. Models now reliably separate benign from malignant tissue, sort tumors into subtypes, grade them, and spot cancer that has spread to the lymph nodes, in some comparisons matching or beating human readers. The more interesting shift is that these tools are no longer just detectors. They are starting to read molecular information directly from a routine stained slide, estimating hormone receptor status, HER2, and proliferation markers that normally require separate and costly immunohistochemistry tests.

What matters most is how this plays out in real practice. When pathologists work with a good AI tool, they tend to become faster and more consistent, and they order fewer follow up stains. That is the honest case for the technology in one sentence. It is not about replacing the pathologist but about making the same person quicker, more accurate, and less reliant on expensive extra testing.

From diagnosis to prognosis

The frontier is no longer just naming what a tumor is, but predicting what it will do. Histology carries subtle signals about outcome that the human eye cannot fully extract, particularly in the tumor microenvironment, the arrangement of immune cells and surrounding tissue around the tumor. About a third of early stage patients relapse within a decade, and many are given aggressive chemotherapy they will never benefit from. The hope is that AI can read these images to sort patients more precisely, sparing some of them treatment they do not need. The prognostic results so far are promising but less settled than the diagnostic ones, which is a fair reflection of how much harder the problem is.

The real bottleneck is data, not algorithms

The most honest theme in the field is that the neural network is rarely what limits success. The data is. Good models need large, diverse, carefully labeled image sets, and those are scarce. Staining and scanning differ from one lab to the next, so a model trained in one place often stumbles in another. Much of the published work leans on a single popular dataset, rarely tests on outside data, and as a result, probably reports accuracy that is a little too flattering. New resources are trying to fix this by offering carefully annotated slides, including the atypical, borderline lesions that are hardest to classify and most in need of better tools. Progress here now depends less on clever architecture and more on building better, shared, well-documented data.

The same caution, from every direction

Whether the work comes from a clinician, an engineer, or a dataset curator, it lands on nearly the same warnings. Models need to prove they generalize beyond the lab that built them. They need to explain their reasoning, because a black box is hard for clinicians to trust or justify to a patient. They need prospective clinical trials rather than just retrospective accuracy. And the questions sitting above the technology, who is responsible for a wrong call, how patient data is protected, and how any of this is regulated, are still unsettled.

The direction of travel is clear enough. The next phase points toward foundation models trained on huge amounts of pathology data, systems that combine the slide with genomic and clinical records, tools that can show their work, and ways of training shared models across institutions without pooling sensitive data.

The takeaway

The field has genuinely reached, and on some specific tasks, passed, expert-level performance, and AI is starting to make working pathologists measurably better rather than merely matching them in a study. What stands between that and everyday clinical use is not a smarter algorithm but the unglamorous groundwork: standardized staining and scanning, bigger and more representative datasets, rigorous validation, models that can explain themselves, and clear rules for accountability. The exciting part is mostly done. The important part is what remains.

References

  1. Jiang, B., Bao, L., He, S., Chen, X., Jin, Z., & Ye, Y. (2024). Deep learning applications in breast cancer histopathological imaging: diagnosis, treatment, and prognosis. Breast Cancer Research, 26, 137. https://doi.org/10.1186/s13058-024-01895-6
  2. LakshmiPriya, C. V., Biju, V. G., Vinod, B. R., & Ramachandran, S. (2024). Deep learning approaches for breast cancer detection in histopathology images: A review. Cancer Biomarkers, 40, 1–25. https://doi.org/10.3233/CBM-230251
  3. Datwani, S., Khan, H., Niazi, M. K. K., Parwani, A. V., & Li, Z. (2025). Artificial intelligence in breast pathology: Overview and recent updates. Human Pathology, 162, 105819. https://doi.org/10.1016/j.humpath.2025.105819
  4. McCaffrey, C., Jahangir, C., Murphy, C., Burke, C., Gallagher, W. M., & Rahman, A. (2024). Artificial intelligence in digital histopathology for predicting patient prognosis and treatment efficacy in breast cancer. Expert Review of Molecular Diagnostics, 24(5), 363–377. https://doi.org/10.1080/14737159.2024.2346545
  5. Brancati, N., Anniciello, A. M., Pati, P., et al. (2022). BRACS: A Dataset for BReAst Carcinoma Subtyping in H&E Histology Images. Database, 2022, baac093. https://doi.org/10.1093/database/baac093
  6. Tafavvoghi, M., Bongo, L. A., Shvetsov, N., Busund, L.-T. R., & Møllersen, K. (2024). Publicly available datasets of breast histopathology H&E whole-slide images: A scoping review. Journal of Pathology Informatics, 15, 100363. https://doi.org/10.1016/j.jpi.2024.100363
  7. Tahir, M., Hu, Y., Kumar, H., et al. (2025). A Comprehensive AI-Based Approach in Classifying Breast Lesions: Focusing on Improving Pathologists' Accuracy and Efficiency. Clinical Breast Cancer, 25(6), e818–e825. https://doi.org/10.1016/j.clbc.2025.03.016
  8. Wang, G., Jia, M., Zhou, Q., et al. (2024). Multi-classification of breast cancer pathology images based on a two-stage hybrid network. Journal of Cancer Research and Clinical Oncology, 150, 505. https://doi.org/10.1007/s00432-024-06002-y

 

Comments

Post a Comment

Popular posts from this blog

Influenza Virus Evolution: Challenges of Antigenic Drift and Shift in Vaccine Design and Response

Image
  I ntroduction The under lying principles of influenza vaccination have not altered significantly over 70 years. The greatest challenge lies in the virus continuously evolving by antigenic drift and shift, reducing the efficacy of the vaccine and requiring ongoing revision. Curren tly, available vaccines rely on the tactic of attacking specific viral proteins, which are too few to defend against all the strains. Developing effective vaccines for influenza is complicated due to the virus’s propensity to mutate frequently, especially in the hemagglutinin (HA) protein. Even though antibodies to the HA head are known to neutralize infection, this region is highly variable, and therefore, vaccines are not very effective in the long term. Thus, vaccines currently in use offer temporary protection and must be updated annually. Influenza A and B viruses continue to cause severe disease across the globe, and evolutio n is tracked by the WHO for vaccine composition. Although this system en...
Read more

How Illumina Sequencing Works

Image
Illumina sequencing is the most widely used DNA sequencing platform today. It powers everything from cancer genomics to microbiome research, including antimicrobial resistance (AMR) surveillance. Overview of the Workflow The entire Illumina sequencing process is divided into four major stages:   Stage Name Purpose 1 Library Preparation Fragment DNA and attach adapter sequences to each fragment end 2 Cluster Generation Amplify each fragment into ~1,000 identical copies on the flow cell 3 Sequencing by Synthesis Add fluorescent nucleotides one at a time and image each cycle 4 Data Analysis Base-call reads and align to a reference genome or assemble de novo   Stage 1 : Li brary Preparation We start with a biological sample like tissue, blood, a swab, or environmental material, and extract genomic DNA. This raw DNA is far too long to sequence directly, so the first step is fragmentation: breaking it into short pieces, typically 150–300 base pairs long. Fragmentation is performed ...
Read more