AI use in breast cancer diagnosis through the detection of tumour biomarkers in serum

The study by Mao et al. (2023) introduces OncoSeek, an AI-driven model trained on seven serum protein tumour biomarkers, including CEA and CA15-3, for multi-cancer early detection9. The conventional method, which relies on individual thresholds for tumour biomarkers, produced high false-positive rates with a specificity of 56.9%. This means that when fixed cutoff values are used to interpret markers like CEA or CA15-3, many people without cancer are incorrectly identified as having the disease. This high false-positive rate highlights the limitations of conventional threshold-based diagnostics and sets the stage for AI models that can interpret complex patterns across multiple markers. OncoSeek significantly improves specificity to 92.9% and accuracy to 84.3%. This is achieved by employing machine learning models such as generalised linear models (GLM), gradient-boosted machines (GBM), and random forests (RF). These models are capable of capturing non-linear marker interactions. Such interactions involve complex, non-additive relationships among biomarkers, where the effect of one marker on diagnosis depends on the presence or concentration of others. This approach offers a significant advantage over fixed cutoff values, which treat each biomarker independently and fail to account for their interdependencies, often resulting in higher rates of false positives or false negatives. Unlike linear models that assume fixed individual contributions, AI models such as GBM and RF can capture these intricate dependencies, therefore improving diagnostic sensitivity and specificity by learning synergistic or conditional patterns across multiple biomarkers9. This demonstrates AI’s potential to aggregate weak individual signals from multiple markers into a clinically robust diagnostic output.

Luan et al. (2023) confirmed the findings of Mao et al. by demonstrating that tumour biomarkers detection assisted by AI can achieve high specificity in multi-cancer early detection. The same OncoSeek platform was used, which processed inputs from seven serum-based protein tumour biomarkers including CEA and CA15-3, which are specific to breast cancer. The remaining 5 markers AFP, CA125, CA19-9, CYFRA21-1, and NSE are associated with other cancers, enabling the model to support multi-cancer detection10. Each patient’s serum profile was transformed into a feature vector containing marker concentrations, which were labelled based on their confirmed cancer status and tissue of origin. During training, the model learned to assign probabilities to multiple cancer types simultaneously, optimising its performance using stratified cross-validation and grid search. Importantly, this approach enabled the model to detect synergistic patterns between markers. For example, elevated CA15-3 was interpreted as more significant when CEA or other markers were concurrently abnormal. The end result was a marked increase in diagnostic accuracy of up to 84.3% and specificity of 92.9%, particularly in distinguishing malignant from benign or healthy cases, compared to fixed-threshold rules10. While Mao et al. (2023) introduced and validated OncoSeek’s diagnostic performance in a large population, Luan et al. (2023) conducted an independent evaluation of the same OncoSeek platform across multiple institutions in China, involving both urban and rural clinical centres. This multi-site validation confirmed that OncoSeek maintained high diagnostic performance with specificity of 92.9% and accuracy of 84.3%, even when tested on diverse patient populations and laboratory environments9,10. The consistency of results across institutions demonstrated the model’s robustness, highlighting its potential for widespread clinical adoption and reliable integration into real-world workflows regardless of institutional variability. Both studies employed the same AI model architecture and seven-marker serum panel, reinforcing the robustness and reproducibility of OncoSeek for multi-cancer early detection.

AI use in breast cancer diagnosis through histopathology

Huang demonstrated that AI models can maintain consistent diagnostic performance even when applied to histological slides stained using different protocols11. This work lays the foundation for evaluating AI resilience across diverse real-world laboratory settings. These stains were primarily H&E and its colour-normalised variants, excluding antigen-based methods such as IHC or immunofluorescence. Their study focused on developing a convolutional neural network (CNN)-based model that is robust to variations in routine H&E staining caused by differences in reagents, laboratory protocols, and microscopes. By employing cross-domain learning strategies and training on multi-stain datasets, the model achieved over 90% classification accuracy in distinguishing between benign and malignant lesions. It was also able to differentiate among histologic subtypes with high accuracy. These results were observed in multi-institution validation sets that included cross-stained H&E variants11. This increases the model’s robustness to real-world staining inconsistencies. However, the study did not explore the model’s interpretability or offer insights into which morphological features were most predictive. This limits its potential utility as a decision support tool for pathologists. Additionally, while classification accuracy was high, the study did not report sensitivity and specificity for each individual staining domain. Without these detailed metrics, it is challenging to identify specific scenarios where the model may fail (failure modes) or determine how reliably it performs across different histologic classes (class-specific reliability). Nonetheless, the model demonstrated strong cross-stain robustness, which supports its generalisability. This capability offers promise for deployment in real-world laboratories where staining inconsistencies are unavoidable. In summary, Huang’s study illustrates a feasible pathway for integrating AI into pathology workflows that rely on H&E staining.

In the work by Sandbank, a CNN-based model was trained to analyse whole slide images (WSIs) of breast biopsies stained with H&E and hematoxylin-eosin-saffron stain (HES), which adds saffron to enhance the contrast of connective tissue12. The algorithm learned to classify 51 morphological and diagnostic classes, including ER/PR/HER2-related subtypes, various grades and ductal carcinoma in situ (DCIS). Training was performed on a dataset of digitised WSIs, with expert-annotated regions of interest. Through supervised learning, the CNN extracted features such as nuclear morphology, tissue architecture, and staining intensity. Model optimisation was performed using backpropagation and stochastic gradient descent to minimise prediction error and refine feature representations relevant to diagnostic classification. The high-resolution images were divided into smaller patches, enabling the model to assign probabilities to diagnostic labels at both patch and whole-slide levels. It achieved AUCs of 0.99 for invasive carcinoma and 0.98 for DCIS across a diverse validation cohort, showing near-human-level accuracy12. Notably, the AI model served as a second-read system to detect misdiagnosed cases, reinforcing its value as a real-time quality control tool. This study exemplifies how well-curated datasets and granular annotations can result in high-performing diagnostic AI models.

In contrast, Malherbe emphasises the underexplored potential of AI in decoding tumour microenvironment interactions, especially the spatial relationships between immune, stromal, and tumour cell populations13. Despite the increasing availability of multiplex IHC and spatial omics technologies, most AI models reduce this complexity into simplified feature vectors or overlook spatial context altogether. For example, models processing multiplexed ion beam imaging (MIBI) and co-detection by indexing (CODEX) data often reduce the spatial organisation of immune and stromal cells into scalar quantities which is single summary values that condense rich, multi-dimensional spatial data into basic metrics, such as the total number of immune cells, the average distance between cells, or the proportion of a specific cell type. These scalar values discard spatial topology, meaning that although the presence of individual cells is captured, the biologically meaningful spatial interactions such as immune cells clustering around tumour cells, which may indicate an active immune response will be lost14,15. This simplification can obscure key spatial patterns that are highly relevant for predicting or interpreting immunotherapy outcomes13. While Malherbe highlighted this limitation and recommended the use of graph neural networks (GNNs) and spatially resolved transcriptomics as possible avenues, these strategies were not formalised into a structured model. Building upon this insight, the present review proposes a conceptual framework (Figure 2) that explicitly integrates GNNs with spatially resolved omics data to capture the biological relevance of cell-cell interactions in breast cancer. This model illustrates how GNNs can preserve spatial topology while incorporating domain-specific biological priors, thereby addressing the limitations of conventional CNN-based models that flatten tissue architecture into simplified statistics. By advancing Malherbe’s suggestion into a schematic pipeline, our framework provides a visual roadmap for how AI could be embedded into immuno-oncology workflows to enhance biomarker interpretation and immunotherapy decision-making.

Choudhury & Perumalla (2021) developed a CNN-based model to classify breast cancer using 50x50 pixel image patches derived from digitised H&E-stained histopathology slides, labelled as invasive ductal carcinoma (IDC+) or non-IDC (IDC-)16. The model architecture included convolution and pooling layers, fully connected layers, and a softmax output for probability generation. Given the dataset’s small size of 5,547 patches, the authors used data augmentation such as image rotation to improve generalisability16. While the model achieved a classification accuracy of 78.4%, it lacked external validation and failed to benchmark against human pathologists. Tumour biomarker integration was also absent. These limitations reduce the model’s applicability in clinical workflows. Compared to Huang’s or Sandbank’s studies, this Choudhury & Perumalla highlighted how dataset size and lack of external benchmarks can constrain AI model performance and translational impact.

AI-based prediction of tumour biomarkers expression and treatment outcome in breast cancer patients using histopathology data

The study by Meshoul et al. (2022) advances the field by predicting biomarker status of ER, PR, and HER2 directly from H&E-stained images, bypassing the need for IHC staining17. By training deep learning models on thousands of H&E-stained slides with biomarker labels verified by corresponding IHC results, the authors developed an interpretable AI framework capable of predicting ER, PR, and HER2 expression. The model employed a CNN trained in a supervised manner, in which image patches were paired with known expression levels and optimised through backpropagation. Through backpropagation, the network learned to identify morphological features such as nuclear size, chromatin texture, and glandular structure that statistically correlated with specific biomarker profiles. Although the model does not detect antigen presence via biochemical interaction, it functions as a surrogate by inferring expression likelihood based solely on visual morphological cues, thus effectively simulating IHC status prediction through image-based inference17. This approach not only reduces reagent use and turnaround time but also demonstrates the feasibility of AI-based histology-to-phenotype prediction as a potential alternative to conventional biomarker assays.

Gamble et al. (2021) extended this approach by applying AI to infer HER2, ER, and PR biomarker status directly from histomorphological features, thereby supporting the emerging concept of digital-only biopsy analysis18. Their CNN-based model was trained using annotated H&E-stained slides from The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA) cohort, with corresponding IHC data serving as the reference standard for biomarker status. The model learned to associate specific histological patterns such as glandular architecture, nuclear pleomorphism, and overall tissue organisation with known biomarker expression profiles. Although the model did not rely on biochemical detection, it effectively simulated IHC scoring by leveraging statistically derived visual cues, thereby offering a cost-effective and scalable strategy for large-scale biomarker profiling. The model achieved AUC values of 0.91 for ER, 0.95 for PR, and 0.93 for HER2 prediction, indicating high discriminative performance. The AUC metric, or area under the receiver operating characteristic (ROC) curve, reflects the model’s ability to distinguish between positive and negative cases across all possible decision thresholds. An AUC of 1.0 denotes perfect classification, while 0.5 indicates performance no better than random guessing18. Thus, the AUC values reported by Gamble et al. suggest that the model achieved excellent discriminative power in classifying biomarker expression status from histomorphological features alone. However, the study did not report sensitivity or specificity metrics, nor did it include a direct comparison between model predictions and conventional IHC scoring. Furthermore, the model’s generalisability was constrained by its reliance on a single dataset (TCGA-BRCA), and external validation across diverse clinical settings was not conducted.

Dodington et al. (2021) demonstrated the capability of AI to evaluate both treatment response and molecular subtype prognosis across distinct categories of breast cancer, including luminal A, luminal B, HER2-enriched, basal-like, and triple-negative subtypes19. Each of these molecular classifications carries specific clinical implications: luminal A tumours are typically hormone receptor-positive with low proliferative activity and favourable prognosis; luminal B tumours exhibit higher proliferation and may require chemotherapy; HER2-enriched tumours overexpress HER2 and benefit from targeted therapies; and basal-like and triple-negative subtypes tend to be more aggressive, lacking hormone receptor and HER2 expression, often resulting in poorer outcomes with limited therapeutic options2. To address this clinical heterogeneity, Dodington and colleagues developed a supervised machine learning model trained on nuclear morphometric features such as nuclear size, shape, and chromatin texture that were extracted from digitised H&E-stained biopsy slides19. Crucially, these slides were labelled with patient-level therapy outcomes, enabling the model to learn associations between pre-treatment histomorphology and pathological complete response to neoadjuvant chemotherapy. The model achieved an F1 score of 0.87, indicating a strong balance between sensitivity and precision in identifying responders while minimising false predictions19. Pathological complete response is defined as the absence of residual invasive cancer in the breast and axillary lymph nodes following treatment (ypT0/is ypN0) and is widely recognised as a surrogate marker for improved long-term survival in aggressive breast cancer subtypes20. This performance significantly outperformed a logistic regression baseline, suggesting the model was better equipped to detect subtle morphological patterns predictive of treatment efficacy. Notably, their analysis revealed that specific nuclear morphologies correlate with therapeutic response, supporting the hypothesis that tumour architecture and cellular characteristics encode predictive cues. However, the study’s limitations include its relatively small retrospective cohort (n = 82), risk of selection bias, and the absence of external validation across multiple institutions.

Building on these developments, Garberis et al. (2025) introduced a deep learning model known as RlapsRisk, designed to predict five-year metastasis-free survival (MFS) in patients with ER+/HER2- early breast cancer using routine HES-stained slides21. The model was trained on a large retrospective cohort (GrandTMA; n = 1429) and externally validated on the prospective CANTO cohort (n = 889). It utilised a multi-stage architecture that incorporated self-supervised learning for feature extraction and a multiple instance learning (MIL) approach to predict outcomes from entire whole-slide images. Notably, RlapsRisk captured prognostic morphological patterns such as high nuclear pleomorphism, mitotic activity, and tumour-stroma spatial interactions without relying on manual annotations by pathologists. The resulting risk score demonstrated independent prognostic value beyond established clinico-pathologic factors and improved discrimination performance when combined with standard clinical models (C-index improved from 0.76 to 0.80 in the validation cohort, p < 0.005)21. Importantly, the model showed robust stratification performance across clinically relevant subgroups including patients with intermediate histological grade or clinical risk and provided interpretable outputs via tile-level Shapley value analysis, supporting clinical trust and transparency. By directly addressing the clinical endpoint of distant recurrence and requiring only standard digital slides, this study exemplifies the real-world potential of AI to enhance prognostic assessment and inform adjuvant treatment decisions, particularly where access to genomic assays remains limited.

Tanveer et al. (2025), Jones et al. (2022), Choudhury & Perumalla (2021), and Nassif et al. (2022) collectively reinforce the central role of CNNs in breast cancer diagnostics, particularly in image-based classification and feature extraction16,22,23,24. However, these studies differ substantially in methodological rigor, dataset scale, and clinical applicability. Some rely on small, retrospective cohorts lacking external validation, while others incorporate more diverse data sources yet fall short in explainability or clinical workflow integration. For instance, Tanveer et al. developed an ensemble machine learning framework that combined radiomic features derived from breast imaging with serum protein biomarker data. The model was trained using a supervised learning strategy in which multimodal features were input into several base classifiers such as support vector machines and random forests and subsequently aggregated by a meta-classifier to improve prediction robustness. Although the model reported an accuracy of 91.2% in cancer detection, it was not externally validated on independent or multicentre datasets, thereby limiting its generalisability22.

Jones proposed a hybrid deep learning pipeline that integrated imaging and clinical metadata, including age, menopausal status, family history, hormone receptor status, tumour size, and comorbidities23. Their fusion architecture combined features extracted from digital mammography with those from electronic health records, using convolutional and fully connected layers. The model was trained on labelled diagnostic outcomes to predict both malignancy risk and molecular subtype23. However, the study omitted key components related to model interpretability such as feature attribution methods or saliency mapping and have not address data governance issues, including consent, data-sharing policies, and deployment logistics in clinical settings.

Taken together, these studies emphasise the technical promise of CNNs but also reveal recurring limitations such as insufficient explainability, lack of prospective or external validation, and limited integration into real-world diagnostic workflows. Critically, few of the models have progressed beyond in silico evaluation to full clinical deployment. As such, their robustness in operational environments where variability in patient demographics, imaging protocols, and institutional infrastructure inevitably remains untested.

Scoping synthesis of findings by data modality, validation design and explainability.

Across the eleven reviewed studies, three major thematic patterns emerge that collectively define the current state of AI applications for tumor biomarker detection in breast cancer: (1) data modality and diagnostic scope, (2) validation and generalisability, and (3) explainability and clinical integration (Table 2).

Explainability and clinical integration.

A further recurring theme is the limited interpretability of deep learning models. With the exception of Meshoul et al. (2022), who applied SHAP-based interpretation, most CNN-based systems operate as “black boxes,” generating predictions without transparent biological or morphological justification. This opacity constrains clinician trust and hinders regulatory approval. Although high AUCs (e.g., 0.91–0.95 for ER/PR/HER2 prediction in Gamble et al., 2021) demonstrate technical performance, the lack of explainable AI mechanisms, such as saliency or attention mapping, continues to impede clinical adoption despite high reported AUCs. Few models have progressed beyond retrospective validation to deployment in pathology or diagnostic laboratories. Regulatory and interoperability barriers further delay translation. Moving forward, embedding saliency mapping, attention-based visualisation, and human-AI collaborative validation into model design could facilitate both interpretability and trustworthiness. Beyond algorithmic transparency, integrating health-economic and human-factor analyses into AI implementation research is essential for sustainable clinical adoption (Table 2).

Together, these thematic insights reveal a dynamic yet fragmented research landscape. The reviewed studies nonetheless highlight the growing potential of AI to transform tumour-biomarker detection for better breast-cancer diagnostics; however, several limitations must be addressed to ensure clinical relevance and long-term integration (Figure 3).

Theme	Key Patterns (Across 11 Studies)	Representative Examples	Main Gaps / Needs
Data Modality & Diagnostic Scope	AI used mainly in serum (OncoSeek), histopathology (CNN-based H&E/HES models) or morphology-to-biomarker prediction; very few true multimodal designs.	Mao 2023; Luan 2023; Sandbank 2022; Gamble 2021; Jones 2022.	Integration across imaging + omics + clinical data remains rare; need for unified multimodal pipelines.
Validation & Generalisability	Most studies internal, single-centre, retrospective; only 3/11 had external or multicentre validation.	Garberis 2025; Luan 2023; Sandbank 2022.	Require large, prospective, multi-institutional validation and consistent reporting of sensitivity, specificity, calibration.
Explainability & Clinical Integration	Only Meshoul 2022 used formal XAI (SHAP); others remain “black-box.” Few clinical deployments.	Meshoul 2022; Gamble 2021; Choudhury 2021.	Need embedded saliency/attention mapping, human-AI review, and integration with regulatory, cost-effectiveness, and workflow studies.