AI Scoring of Immune Cells in Breast Cancer Tissue Is Booming, but Almost No Model Is Ready for the Clinic

Machine learning models that read immune cells on breast cancer slides are multiplying fast, but a new systematic review concludes that the field is technically active yet clinically under-validated. The analysis, published in Frontiers in Oncology, mapped 51 histopathology-image studies that use machine learning to assess tumor-infiltrating lymphocytes, or TILs, and found that while external validation and clinical benchmarking rise as studies move closer to patient-level decisions, calibration and formal clinical-utility analysis are almost entirely absent. Calibration was reported in just one of 18 prognostication studies and in none of the detection or treatment-response studies, and decision-curve analysis appeared in only that same single prognostication paper.

TILs are immune cells that infiltrate tumors, and their abundance in breast cancer tissue carries real clinical weight. High TIL scores are associated with better outcomes and with stronger responses to neoadjuvant chemotherapy, particularly in HER2-positive and triple-negative breast cancer, while the presence of the PD-1-positive TIL subtype has been linked to poor prognosis. Yet manual scoring by pathologists is subjective and time-consuming, producing variability that limits clinical use despite standardized recommendations from the International TILs Working Group. Automated assessment promises more objective, reproducible quantification from routine stained slides, which is why the computational literature has grown so quickly.

The review team, led by researchers at Khoja Akhmet Yassawi International Kazakh-Turkish University and Al-Farabi Kazakh National University, searched PubMed, Web of Science, Scopus, ScienceDirect, and Google Scholar up to 20 April 2026, registered the protocol prospectively in PROSPERO, and followed PRISMA 2020 reporting guidelines. From 346 retrieved records they initially included 36 studies, then added backward citation tracing of 1,647 references and a targeted Google search, ultimately yielding 51 studies. These were divided into three application clusters defined in advance: 19 detection or classification studies, 18 outcome prognostication studies, and 14 treatment-response prediction studies. Because endpoints and metrics differed so widely across tasks, the authors pooled nothing quantitatively and instead built a descriptive evidence map.

The evidence map revealed a clear translational gradient. External validation was reported in 31.6 percent of detection and classification studies, 61.1 percent of prognostication studies, and 57.1 percent of treatment-response studies. Clinical or pathologist comparator reporting climbed even more steeply, from 63.2 percent in the detection cluster to 94.4 percent in prognostication and 92.9 percent in treatment-response prediction. In other words, studies aimed at patient-level outcomes more often asked whether machine-derived TIL measures add anything beyond what pathologists and clinicopathological variables already provide. Incremental-value analysis followed the same pattern, reported in 16 percent of detection studies, 35 percent of prognostication studies, and 64 percent of treatment-response studies.

But clinical proximity did not equal clinical readiness. Beyond the near-total absence of calibration and decision-curve analysis, the review found that reproducibility remained incomplete everywhere. Code was available for 47.4 percent of detection studies, 38.9 percent of prognostication studies, and 42.9 percent of treatment-response studies. Model weights were available in 33 percent of applicable detection studies, none of the applicable prognostication studies, and 7 percent of treatment-response studies. Data availability, partly inflated because use of named public datasets counted as availability, stood at 63 percent, 50 percent, and 29 percent respectively. The authors caution that access to a public source cohort does not mean the exact analytic dataset, preprocessing workflow, train-test split, annotations, or derived features can actually be reproduced.

Methodological challenges turned out to be structural features of the evidence base rather than isolated flaws. Class imbalance was nearly universal, identified in 94.7 percent of detection studies and 100 percent of both other clusters. Pathologist ground-truth variability was flagged in 78.9 percent of detection and 77.8 percent of prognostication studies but only 42.9 percent of treatment-response studies, likely because response models more often relied on patient-level outcomes than on pathologist-assigned labels. Scanner and stain variability, which directly perturbs the color and texture features many pipelines use to distinguish lymphocyte nuclei from stromal or tumor cells, was addressed or mitigated in 63.2 to 83.3 percent of studies depending on cluster. Fairness or subgroup analysis, by contrast, appeared in only 27.8 to 57.1 percent of studies.

Model families differed sharply by task. Detection and classification studies showed the greatest architectural diversity, with segmentation or object-detection approaches and classical handcrafted machine learning each used in five studies, convolutional neural networks in four, and small numbers of weakly supervised, self-supervised, multimodal, transformer, and benchmark-comparison designs. Prognostication studies concentrated on classical machine learning, segmentation, and CNNs, while treatment-response prediction was the most concentrated cluster of all, using classical approaches in seven studies, CNNs in five, and segmentation in two, with no foundation models, self-supervised methods, or multiple-instance learning at all. Weakly supervised and foundation-model approaches, despite their momentum in the broader computational pathology field, remain barely represented in the TIL-specific literature.

The evidence base also rests on a surprisingly narrow foundation. TCGA-BRCA was the most reused cohort, contributing to nine study rows, followed by the TIGER challenge dataset in six studies and BCSS in four, with METABRIC in three and ICGC and FinHER in two each. The authors warn that repeated analyses of the same underlying patient populations do not constitute independent clinical evidence and may amplify dataset-specific characteristics, meaning the apparent breadth of published studies can overstate the true diversity of represented populations, institutions, and clinical settings. Single-center limitations were identified in 27.8 to 36.8 percent of studies, and small-sample concerns ranged from 22.2 to 57.1 percent, with treatment-response studies particularly affected.

Quality appraisal reinforced the picture. Under the TRIPOD+AI checklist, most studies reported objectives, data sources, model development, and performance metrics well, but outcome definition, clinical interpretation of performance, and implementation items were frequently missing, with many papers reporting technical metrics such as AUC or F1 score without contextualizing clinical relevance. Under PROBAST, every one of the 51 studies was judged at high overall risk of bias, driven mainly by the analysis and outcome domains, including small samples relative to model complexity, lack of external validation, potential overfitting, absent calibration, and weakly standardized outcome labels. The participants and predictors domains, by contrast, showed predominantly low risk.

The review also identified a conspicuous immunotherapy gap: no mature body of machine learning models predicting response to immune-checkpoint inhibition in breast cancer emerged from the evidence map, even though TILs are biologically central to anti-tumor immunity and especially relevant in triple-negative disease. The authors attribute this to smaller and more heterogeneous immunotherapy trials, immune-related response patterns such as pseudo-progression that complicate ground-truth labeling, and the absence of a large annotated multi-institutional immunotherapy-response TIL cohort. Their prescription is cluster-specific rather than generic: future studies should predefine intended use, report annotation and interobserver agreement, distinguish stromal from intratumoral compartments, validate on genuinely independent cohorts, report calibration and decision-curve analysis, and release code, weights, and study-specific data. None of the included studies addressed regulatory pathways or workflow integration, which the authors say confirms that deployment readiness, not further model development, is the field’s principal unmet need.

Subject of Research: Machine learning-based assessment of tumor-infiltrating lymphocytes in breast cancer histopathology and its translational readiness

Article Title: Machine learning-based assessment of tumor-infiltrating lymphocytes in breast cancer histopathology: a systematic review and evidence-gap analysis of translational readiness

Article References: Dunenova, G., Sarsembayev, A., Oshibayeva, A., Kyrykbayeva, S., Abdrakhmanova, Z., Omarova, B., Ivankov, A., Ualiyeva, A., Zhakypova, B., Zhagiparova, Z., Nuskabayeva, G., Suleimenova, M., & Glushkova, N. (2026). Machine learning-based assessment of tumor-infiltrating lymphocytes in breast cancer histopathology: a systematic review and evidence-gap analysis of translational readiness. Frontiers in Oncology, 16, Article 1908349. https://doi.org/10.3389/fonc.2026.1908349

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1908349

Keywords: breast cancer, tumor-infiltrating lymphocytes, machine learning, computational pathology, systematic review, external validation, calibration, decision-curve analysis, reproducibility, treatment-response prediction, prognostication, evidence-gap analysis