Can breast cancer AI tools reliably detect male breast cancer?

The sharpest answer to “Can breast cancer AI tools reliably detect or support male breast cancer care?” does not come from a slogan about male breast cancer awareness. It comes from a performance drop: an attention-based multiple instance learning model trained to predict estrogen receptor and progesterone receptor status from H&E whole-slide images reached an AUROC of 0.86 in female breast cancer, then fell to 0.66 when applied to male breast cancer.[1]

That is not a cosmetic decline. In a biomarker prediction task, AUROC is a measure of how well the model separates cases with and without the target status across decision thresholds. A model at 0.86 may look promising enough to motivate clinical development, triage research, or workflow studies. A model at 0.66 is much closer to a weak discriminator. If the patient in front of the pathologist is male, the overall female performance number is not the number that matters.

The pathology result that should make deployment teams pause

Chatterji and colleagues tested whether prediction models for hormone receptor status in female breast cancer extend to male breast cancer. The task was specific: use digitized H&E-stained whole-slide images to predict ER and PR status. The modeling approach was not a crude image classifier bolted onto a vague clinical endpoint; it used attention-based multiple instance learning, a design that can learn from slide-level labels while attending to informative regions within large pathology images.[1]

That matters because H&E-to-biomarker prediction is one of the more interesting promises in computational pathology. If it works, a model may help identify patterns in routine morphology that correlate with molecular or immunohistochemical states. It does not replace validated receptor testing, but it can be studied as a support tool, a quality-control signal, or a way to prioritize additional testing in settings where pathology resources are constrained.

Hormone receptor status is also not a peripheral issue in male breast cancer. More than 90% of male breast cancers are ER-positive, which means an AI failure around hormone receptor prediction lands near the center of how these tumors are classified and treated, not at the edge of an academic benchmark.[1]

The model’s female performance, taken alone, could be read as encouraging. The male result changes the interpretation. The same pipeline that separated receptor status well in female breast cancer did not generalize well to male breast cancer. The study authors did not reduce that failure to the convenient explanation that male cases are simply rare and therefore underpowered. Their interpretation points toward sex-specific morphology: male breast cancer may contain tissue patterns relevant to receptor prediction that a model trained on female disease does not learn adequately.[1]

That distinction is important. Small sample size is a real obstacle in male breast cancer AI, and no serious dataset plan can wish it away. But if the problem is also biological and morphological, then adding a few male cases as an afterthought is not validation. It is decoration. The validation population has to be visible, large enough to be interpretable, and analyzed separately enough that failure is not hidden inside the dominant group.

Why this is not just a rare-disease footnote

Male breast cancer is rare, but rare is not the same as clinically negligible. The CDC estimates about 2,670 new cases of breast cancer in men in the United States in 2026, with about 530 deaths each year.[2] Breastcancer.org, citing American Cancer Society data, describes a rise in U.S. male breast cancer cases from about 1,400 in 2000 to about 2,800 in 2025, and reports a 272% global increase from 1990 to 2021.[3]

The mortality context also matters. Chatterji and colleagues note that male breast cancer has higher mortality than female breast cancer, including a reported 19% higher mortality in men after adjustment.[1] That does not mean an AI model caused the disparity, and it does not mean every AI tool will worsen it. It does mean that a population already facing worse outcomes should not be asked to accept unreported subgroup performance as a matter of administrative convenience.

The everyday clinical pathway is already shaped by gendered assumptions. Men may not recognize a breast symptom as potentially malignant, and clinicians may not immediately place breast cancer high on the differential when a male patient presents with a lump, nipple change, or discharge. AMWA describes how stigma and gendered assumptions can contribute to delayed diagnosis in men.[4] In that setting, a weakly validated AI output is not an abstract fairness problem. It becomes one more place where the patient’s sex can quietly determine the reliability of the system around him.

This is the point where broad male breast cancer awareness becomes clinically specific. Awareness is useful if it changes who gets examined, who gets biopsied, and whose data are considered necessary for evidence. It is not enough if the AI evidence package still treats male patients as too few to analyze and too inconvenient to label.

The FDA-cleared tool problem is evidence visibility, not proof of universal failure

The Chatterji study directly tests a pathology model for hormone receptor prediction. It does not directly test mammography computer-aided detection, breast density tools, image triage systems, or risk prediction algorithms in men. That boundary should be kept intact. A pathology failure is not automatically a mammography failure.

The harder problem is that current published evidence for FDA-cleared breast cancer AI tools does not give clinicians the sex-stratified male performance data they would need to know. Across cleared breast imaging and pathology AI products, public evidence has generally emphasized overall performance, lesion-level detection, reader-assist outcomes, or workflow claims. Male-specific performance is not presented in a way that lets a radiology group, pathology department, or governance committee answer the basic deployment question: what happened when this tool was used on male breast cancer cases?

That absence should not be inflated into a claim that every cleared tool is unsafe for men. It is better described as an evidence-visibility failure. One rigorous study in a related breast cancer AI task shows that sex-specific generalization can fail badly. The public evidence for deployed tools does not show whether the same kind of failure exists, is smaller, or has been mitigated. For a clinical user, those three possibilities are very different. The documents should let users distinguish them.

This is especially relevant for mammography AI because most breast imaging AI evidence is built around screening populations, and routine breast screening is overwhelmingly organized around women. Men are usually diagnosed through symptom-driven evaluation, not population screening. A model trained and tested in a female screening environment may still be useful in some male diagnostic contexts, but that is a hypothesis requiring evidence, not a property inherited from the female cohort.

What the AUROC gap changes in practice

A drop from 0.86 to 0.66 changes how a model should be allowed to appear in a clinical workflow. At 0.86, a team might consider prospective evaluation as a decision-support signal, provided the use case is narrow and the standard test remains in place. At 0.66, the model should not be presented to clinicians as if its female validation carries over. The output may still be studied, but it belongs behind a warning label, not inside a silent inference pipeline.

The risk is not only a wrong answer. It is a wrong answer with institutional confidence around it. A pathologist or oncologist can often recognize when a laboratory result conflicts with the rest of the case. An AI score is different when its subgroup uncertainty is invisible: the clinician sees a number, a heatmap, a binary flag, or a risk category without seeing the population boundary that produced it.

For male breast cancer, the likely points of failure are not limited to the final prediction. Bias can enter when cases are collected, when slides or images are labeled, when training batches are sampled, when performance is averaged, and when the product label describes intended use. If male cases are too few to support a precise estimate, that uncertainty should be stated. If there are no male cases in the validation set, that should be stated even more plainly.

The responsible answer is narrower than the marketing claim

For the best-studied pathology example, the answer is no: a model trained on female breast cancer data did not reliably generalize to male breast cancer for ER/PR prediction from H&E whole-slide images.[1] For mammography CAD systems and risk prediction tools, the answer is not that failure has been proven in the same way. The answer is that male-specific evidence is largely absent from public validation, and the pathology result makes that absence harder to excuse.

A useful evidence standard would not require every developer to solve rare-disease AI before releasing any breast cancer model. It would require developers and regulators to stop letting aggregate performance do the work of subgroup validation. If the intended-use population includes men, the validation report should say how many male cases were included, what the model did on them, how uncertain that estimate is, and whether the model’s label limits use when male-specific evidence is insufficient.

Dataset design also has to become more deliberate. Male breast cancer cases may need multi-institutional pooling, registry-linked slide and imaging collections, federated evaluation, or prospective enrichment rather than passive inclusion. None of those strategies is simple. But the alternative is simpler only on paper: build the model where data are abundant, publish a clean overall number, and discover the boundary when the patient belongs to the population the model did not learn.

Male breast cancer is too small a population to be treated casually and too clinically distinct in the available evidence to be silently absorbed into female-trained AI performance claims. The honest deployment position is sex-stratified validation, transparent labeling, and visible uncertainty before the tool reaches the clinician trying to care for the man in front of them.

References

1. Prediction models for hormone receptor status in female breast cancer do not extend to males — npj Breast Cancer, 2023
2. About Breast Cancer in Men — CDC
3. Male Breast Cancer Cases Are Increasing Worldwide — Breastcancer.org
4. Breast Cancer in Men: How Stigma and Gendered Assumptions Delay Diagnosis — American Medical Women’s Association