How artificial intelligence applied to digital pathology could guide treatment personalization in breast cancer

Abstract

Introduction

The global increase in breast cancer incidence is determining substantial augment of pathologists’ workload. Digital pathology (DP), which consists of digitizing haematoxylin–eosin (H&E)-stained slides into whole-slide images, enables new workflows and horizons. The integration of artificial intelligence (AI) with DP led to the emergence of computational pathology (CP), which has the potential of offering valuable diagnostic, prognostic and predictive tools. The aim of this narrative review is to provide an overview of the state of the art in this rapidly evolving field.

Diagnostic, predictive and prognostic application

Several AI tools demonstrated enhanced performances as compared with visual evaluation of pathologists, such as identification of invasive breast cancers, sentinel lymph node metastasis, histological grading, Ki-67, estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor 2 (HER2) and programmed death-ligand 1 (PD-L1) quantification. Due to this, most of them are currently authorized in the European market. Beyond diagnosis, CP has shown strong potential in predicting, from a simple H&E digital slide, treatment response (e.g. neoadjuvant chemotherapy, anti-HER2 therapy and endocrine therapy), immunohistochemistry status (ER, PR, HER2, PD-L1), molecular alterations (e.g. BRCA1/2 and homologous recombination deficiency status), genomic signatures (e.g. Oncotype DX) and ultimately the patient’s risk of recurrence.

Conclusion

Despite challenges such as high validation costs, initial economic investments and potential AI biases, CP holds great promise. Continued effort to address these barriers and improve AI reliability are crucial. Emerging AI tools, including agent-driven systems, point towards the potential integration of CP into routine clinical workflow practice, even though clinical validation has not yet been established.

Highlights

• •CP combines DP and AI.
• •AI tools outperform traditional pathology in diagnostic tasks.
• •This approach predicts treatment response and molecular traits from stained slides.
• •Several tools using AI are approved for clinical use in Europe.
• •Integration into routine workflows needs more validation and implementation.

Introduction

Background

Breast cancer diagnosis is increasing worldwide, with 2.3 million diagnoses in 2022 and 670 000 deaths, and projections estimating 3.2 million diagnoses and >1 million deaths in 2050.¹

The increasing number of diagnoses is reflected as a rise in workload for pathologists, representing an health organization problem, given the worldwide shortage of pathologists.²

In parallel, several molecular biomarkers have been integrated into clinical practice, further increasing the burden on pathology departments.³

However, there is still a lack of prognostic and predictive biomarkers beyond conventional clinicopathological parameters capable of assisting treatment decisions, and some new ones, such as genomic assays, are not always accessible or cost-effective.⁴

Digital pathology (DP) has been developed precisely to address the previously mentioned challenges, such as the increasing workload for pathologists and the lack of reliable biomarkers.⁴ Often integrated with artificial intelligence (AI) applications, this has led to the emergence of the so-called computational pathology (CP).

The aim of this narrative review is to provide an overview of the state of the art in this rapidly evolving field, focusing on the evidence of CP in breast cancer.

Digital pathology and AI

Digital pathology refers to the practice of digitizing haematoxylin–eosin (H&E)-stained glass slides, creating digital images known as whole-slide images (WSIs), by going through additional steps as compared with conventional pathology, such as slide scanning, image management and storage (Figure 1).⁴ This new approach opened a range of new opportunities such as digital collaboration between pathologists, the possibility of remote working, enhanced teaching and integration with electronic workflows.⁴

Figure 1.

Workflow of digital pathology in oncology. A tissue sample is processed for haematoxylin–eosin (H&E)-stained slide preparation, followed by whole-slide scanning to generate digital whole-slide image (WSI). This WSI may be subsequently analysed using artificial intelligence tools and ultimately approved by a pathologist. AI, artificial intelligence.

Moreover, the digitalization of H&E-stained slides has led to an effort to develop AI tools applied to WSI, until then unavailable for routine activity of pathologists but already used in the research field, defining the conception of CP.⁵

Machine learning (ML) is a subset of AI that enables systems to learn patterns from data and to improve performance without being explicitly programmed.⁶

Deep learning (DL) is a sub-area of ML that utilizes artificial neural networks with multiple layers. When considering CP, convolutional neural networks (CNNs) are the most used neural networks, consisting of layers of artificial neurons, capable of doing complex calculations all taken together, typically designed for image analysis using convolutional techniques.⁷

We can divide CP applications into ‘diagnostic’ and ‘predictive and prognostic’, according to the goal of the tool (Figure 2).

Figure 2.

Applications of computational pathology in breast cancer. Diagnostic applications include identification of invasive cancer, tumour grading, quantification of Ki-67, assessment of ER, PR, HER2 and PD-L1 expression. Predictive applications involve the estimation of treatment response directly from H&E, prediction of IHC profiles, genomic assay results and mutational data. Prognostic applications rely on computational models that stratify patients into high- and low-risk categories, often represented by survival curves. ER, estrogen receptor; H&E, haematoxylin–eosin; HER2, human epidermal growth factor 2; IHC, immunohistochemistry; PD-L1, programmed death-ligand 1; PR, progesterone receptor.

Diagnostic AI tools assist pathologists focusing on features already assessed in clinical practice, thus reducing workload, interobserver variability and turnaround time, with accuracy often reported as area under the curve (AUC) as compared with expert-annotated ‘ground truth’.

Predictive and prognostic tasks involve applications that allow the identification of features going beyond pathologists’ perception, such as prediction of immunohistochemistry (IHC) results directly from H&E slides and novel histological patterns linked to prognostic information, treatment response and genomic and genetic insights.

Diagnostic application of digital pathology

Quality control and cellularity assessment

Tumour quality control and estimation of tumour cellularity are among the most basic yet critical tasks in everyday routine pathology, as they determine which tumour block is selected for clinical trial inclusion and downstream molecular biology testing. Despite its apparent simplicity, manual estimation of tumour cellularity and percentage of necrosis is time-consuming and shows substantial interobserver variability among pathologists.⁸ Simple tissue segmentation models that delineate tumour, non-tumour and necrotic compartments on WSIs can provide objective and reproducible cellularity read-outs to guide selection of the most representative tumour block.^9,10

Breast cancer detection

One of the first applications of AI in DP for breast cancer was invasive breast cancer detection in primary tumour.11, 12, 13 Moreover, several studies assessed sentinel lymph nodes (SLNs), which is a time-consuming and repetitive part of pathologists’ workload.14, 15, 16 Numerous commercial players are now selling Conformité Européenne (CE)-marked products to assess tumour detection (see Table 1).

Table 1.

Non-exhaustive list of commercial companies developing models for breast cancer and biomarker diagnosis according to the item evaluated

Company	DCIS/invasive cancer	Histological type	Grading	Ki67	Hormone receptors	HER2 status	SLN metastasis
Ibex®	✓	✓				✓∗
Paige®	✓ (only invasive)		✓			✓	✓
Primaa®	✓		✓				✓
Ayosin®			✓
Aiforia®			✓	✓	✓	✓	✓
Mindpeak®	✓∗			✓	✓	✓	✓
HaloAI®				✓∗	✓∗	✓∗
Visiopharm®	✓			✓	✓	✓	✓
DeepPath®							✓

With ✓∗, companion developed but not already approved and available for the European market (Conformité Européenne – invitro diagnostic marked) as per the ‘Products—Health AI Register’ website up to 19 September 2025. To date, none of the commercial partners listed have been FDA approved or cleared computational pathology tools specifically for breast cancer.

DCIS, ductal carcinoma in situ; FDA, Food and Drug Administration; HER2, human epidermal growth factor 2; SLN, sentinel lymph node.

A CNN was validated on ∼200 cases from The Cancer Genome Atlas dataset to distinguish breast cancer and non-malignant lesions, resulting in a positive predictive value (PPV) of almost 70% and a negative predictive value (NPV) of 97% as compared with manual evaluation.¹¹ Simultaneously, another CNN was developed in order to retrospectively classify the breast on 221 WSIs, reaching an AUC of 0.96 in discriminating between malignant and benign lesions.¹²

Sandbank et al. and Ibex Company® developed an algorithm to distinguish between invasive and non-invasive breast cancer.¹³ The algorithm, validated on 436 patients, demonstrated high efficacy in discriminating invasive forms and ductal versus lobular subtype, with an AUC of 0.99 and 0.97, respectively.¹³

The CAMELYON16 trial evaluated the efficacy of various DL algorithms in detecting pathological SLNs in 129 WSIs, evaluated also by pathologists with and without time constraint.¹⁴ The best DL models achieved AUC scores similar to those of pathologists without time restrictions, but these top-performing models exceeded the accuracy of pathologists under time constraints.¹⁴

Another retrospective analysis investigated the capability of an AI algorithm to detect lymph node metastasis (macro-, micro-metastasis and isolated tumour cells) in H&E-stained WSIs.¹⁵ The algorithm demonstrated a sensitivity of 100% and an NPV of 100% even for lobular carcinoma.¹⁵

Retamero et al. evaluated the efficacy of another AI algorithm to assist pathologists in 167 WSI SLN analysis.¹⁶ A reduction of nearly 50% in time spent per slide was observed when using AI, and two pathologists showed an increase in sensitivity of 15%-20%.¹⁶

These algorithms could be used as screening tools to reduce unnecessary IHC on SLNs, lowering workload and costs by limiting IHC only to H&E-negative cases.

Tumour grading and mitoses detection

Breast cancer histological grading is one of the determinants of patient prognosis as it reflects the degree of tumour differentiation.¹⁷

The mostly used grading system in breast cancer is the Elston and Ellis (EE), which combines tubule formation, nuclear atypia and mitotic count, the latter being the strongest prognostic indicator.^17,18 However, the process of assessing mitotic count is subjective and time-consuming for pathologists.

Pantanowitz et al. tried to reduce the interobserver variability between pathologists in evaluating mitotic count with the help of AI.¹⁹ Pathologists’ performance was evaluated with or without AI aid, and it was observed that almost 90% of pathologists achieved significantly higher accuracy with AI support, spending on average of 28% less time.¹⁹

Simmat et al. evaluated the accuracy of mitotic scoring on 50 breast cancer cases with and without AI assistance.²⁰ The automatic identification of hot-spots improved the mitotic score accuracy by 15%, as well as reproducibility between pathologists.

Two other studies directly assessed histological grading rather than focusing solely on mitotic count.^21,22 In the first study, validated on ∼600 cases, an AUC of 0.83 was observed, with predictions biologically correlated to cell division pathway.²¹ The second study developed separate DL systems for each EE component, achieving performance comparable to pathologists.²² See Table 1 for commercial players involved in grading evaluation.

IHC evaluation: Ki-67, ER, PR, HER2 and PD-L1

Ki-67

The Ki-67 index is quantified on IHC and gives a proxy of cell proliferation, thus being prognostic and predictive.²³ Regrettably, Ki-67 is not used worldwide due to the lack of reproducibility between pathologists, the high intratumoural heterogeneity and un unclear cut-off value.²³ For these reasons, studies were conducted to evaluate the performance of CP in Ki-67 assessment, and showed that AI algorithms can offer equal or superior consistency compared with manual assessments.24, 25, 26

Rewcastle et al. confronted the reproducibility and the agreement of Ki-67 evaluation of different AI algorithms versus manual assessment on ∼365 patients with luminal-like breast cancer.²⁴ In the study, all scoring methods showed superimposable agreement with pathologists.

Similarly, Kwon et al. investigated the performance of Ki-67 AI evaluation on almost 1000 breast cancers, demonstrating high correlation between visual evaluation and AI evaluation, with the largest divergence occurring in cases with Ki-67 values between 10% and 20%.²⁵

Zwager et al. demonstrated the efficacy of AI algorithm in scoring Ki-67 expression, with automated analysis strongly correlated with manual assessment.²⁶ See Table 1 for commercial players involved in measuring Ki-67.

Estrogen and progesterone receptor

Estrogen receptor (ER) and progesterone receptor (PR) are evaluated manually by estimating positively stained tumour cells on IHC. This evaluation is subjective and varies between observers, especially when ER expression is low (<10%) and staining differences are subtle, determining the increasing interest for AI tools involved in hormone receptor (HR) quantification.²⁷ Interestingly, an ongoing European study coordinated by ‘Karolinska Institute’ aims to evaluate the reproducibility of ER-low scoring by comparing pathologist-read and AI-based scoring systems.

To address these challenges, a study validated a novel AI-based system that measures ER expression, comparing it with visual scoring by pathologists.²⁸ Results demonstrated perfect correlation between automated and manual scoring using a 1% positivity cut-off, confirming the system’s ability to provide a consistent ER evaluation.²⁸

Another study evaluated an automated ER assessment on tissue microarrays, showing good agreement and concordance with manual scoring.²⁹

Moreover, another AI tool was evaluated on ∼200 breast cancer slides to quantify ER and PR.³⁰ Ten pathologists scored the slides with and without AI assistance. The AI results were confirmed by most pathologists (∼90% of cases) and interobserver agreement slightly improved with AI support.

Shafi et al. evaluated another algorithm, which automates the quantification of ER from WSI. Tested on ∼100 invasive breast cancers, the algorithm showed excellent agreement with manual pathologist (∼94%), with few mismatches attributed to technical artifacts.³¹ See Table 1 for companies involved in HR quantification.

HER2

Human epidermal growth factor receptor 2 (HER2) expression is a major biomarker in breast cancer. However, high interobserver variability in HER2 scoring, mostly due to atypical IHC staining and tumour heterogeneity, often complicates the process evaluation.^32,33 Accurate HER2 assessment is particularly important with anti-HER2 therapies such as trastuzumab deruxtecan, where HER2 status orients therapeutic decisions, given the emergence of the so-called HER2-low and ultra-low status, the latter defined as HER2 IHC score greater than no staining (>0) but <1+, including tumours with incomplete or faint/barely perceptible staining in ≤10% of tumour cells.34, 35, 36, 37, 38

This shift raises misdiagnosis risks, as IHC-based HER2 evaluation was initially designed to distinguish amplified versus non-amplified tumours, and is not sensitive enough to differentiate all HER2-negative cancers, so that AI tools may be useful in this setting.

A CNN was used to classify HER2 expression categories according to IHC in 71 breast cancer patients.³⁹ The model demonstrated almost 85% concordance with pathologist assessments and highlighted discordant cases, potentially improving clinical decision-making.³⁹

Another recent study evaluated the performance of a DL model for assessing HER2 status, showing higher concordance with the ground truth as compared with the average agreement among the pathologists.⁴⁰ When focusing on HER2-low tumours, the model identified correctly 95% cases in the validation cohorts, representing a potential aid in this challenging evaluation.

Krishnamurthy et al. aimed to validate a fully automated AI solution for HER2 IHC scoring.⁴¹ The study involved 120 HER2-positive WSIs, and AI demonstrated a strong performance achieving 92% agreement. Moreover, pathologists using AI showed significant improvements in interobserver agreement and accuracy, especially in distinguishing HER2 0+ from 1+ cases.⁴¹ Table 1 summarizes the key commercial competitor.

PD-L1 evaluation

The evaluation of programmed death-ligand 1 (PD-L1) expression on IHC plays a key role in the treatment of advanced triple-negative breast cancer (TNBC), with the possibility of adding immunotherapy to chemotherapy in patients with expression of PD-L1.^42,43 Multiple PD-L1 assays are available, each using different scoring methods and validated with distinct immunotherapies, such as the 22C3 assay and combined positive score (CPS) for pembrolizumab, and the SP142 assay and immune cell (IC) scoring for atezolizumab. For both assays, inter-pathologist agreement, reproducibility and accuracy are moderate both within the same assay and when comparing different assays.^44,45 Unfortunately, AI tools for PD-L1 assessment in breast cancer are limited, with most tools approved in lung cancer.⁴⁶

A DL tool was applied to improve the consistency and accuracy of PD-L1 IC scoring using the SP142 assay.⁴⁷ Using only conventional assessment, inter-pathologist concordance was moderate (∼70%), while employing AI concordance raised to 95%, with accuracy improved both in distinguishing positive versus negative cases and in assessments of varying expression levels.

A study focused on enhancing the precision of evaluating the PD-L1 CPS in TNBC, using a DL-based model.⁴⁸ The study involved ∼20 pathologists, who assessed TNBC samples using both manual methods and AI assistance. When aided by AI, the differences highlighted between pathologists using manual methods were overcome, as an increase in accuracy was seen, reaching 92%.

Predictive, prognostic and theragnostic applications of digital pathology

Predictive data

Prediction of treatment outcome

Neoadjuvant chemotherapy (NAC) is widely adopted for locally advanced or high-risk breast cancers, allowing the assessment of tumour response on surgical specimen.⁴⁹ Achieving a pathological complete response (pCR) after NAC is linked with improved prognosis, especially in TNBC and HER2-positive subtypes.⁵⁰ However, 40%-50% of patients do not achieve a pCR, facing possible insufficient or unnecessary treatment-related toxicities.⁵⁰ Because clinicopathological factors often fail to accurately predict individual risk of recurrence, several studies explored AI-based prediction of response on WSIs.

Saednia et al. aimed to predict patients’ response to NAC from ∼150 breast cancer biopsies using DL algorithm.⁵¹ Among five features extracted from WSIs, graph-based and wavelet features showed the best predictive power. A model based on these items without integrating clinical data reached an AUC of 0.90 in predicting response to NAC, with ∼85% sensitivity and specificity in the validation set (n = 38), surpassing the model based only on clinical data.⁵¹

Another study evaluated the prediction of pCR in 58 patients with HER2-positive or TNBC undergoing NAC, using a DL approach.⁵² Patients who achieved pCR showed higher nuclear intensity, and fewer multifocal tumours, independently associated with treatment response, with almost 80% of cases correctly classified (pCR and non-pCR).⁵²

Focusing on predicting response to NAC in TNBC patients, Du Terrail et al. and Owkin® developed a model using federated learning on WSIs, a method where different AI tools are trained on data stored in different hospitals to subsequently develop a collaborative model.⁵³ The federated learning model achieved an average AUC of ∼0.66, outperforming clinical baseline models (around 0.63 AUC), suggesting new potential markers (e.g. tumour fibrosis).⁵³

Moreover, a DL model was developed to predict pCR from H&E-stained biopsy WSIs of 540 breast cancer patients undergoing NAC.⁵⁴ The model generated a pCR score by identifying tumour epithelium features, with an AUC of 0.84.

Focusing on HER2-positive tumours, few studies have leveraged AI to improve the prediction of HER2 status and response to anti-HER-2 treatment.^55,56 A CNN trained on 187 WSIs predicted trastuzumab response with an AUC of 0.80.⁵⁵ Another approach developed a morphology-based ‘H&E-ERBB2 score’ using a supervised DL model, subsequently tested on >700 patients.⁵⁶ This score correlated with ERBB2 amplification and distant disease-free survival in trastuzumab-treated patients [hazard ratio 0.37, 95% confidence interval (CI) 0.15-0.93, P = 0.034].⁵⁶

Moreover, recent studies highlight that predicting treatment response in breast cancer can potentially include endocrine therapy.57, 58, 59 One of the mechanisms of endocrine resistance in ER-positive tumours is the epithelial-to-mesenchymal transition (EMT), which is hard to identify morphologically.⁵⁷ Using DL on WSIs, models successfully classified EMT phenotypes and predicted endocrine therapy resistance with >80% accuracy.⁵⁸ Additionally, another AI approach tried to predict endocrine therapy response from WSIs by evaluating ER signalling activity (RNA sequencing) beyond the ER receptor, rather than quantifying receptor expression only. The study demonstrated a correlation between higher ER signalling activity with endocrine response and subsequently better progression-free survival (PFS).⁵⁹

Prediction of protein expression omitting IHC

Even though IHC remains the gold standard for evaluating ER, PR, PD-L1 and HER2 expression in breast cancer (the latter with the help of in situ hybridization), they can be time-consuming and can sometimes deplete the material of small core-needle biopsies.^32,60, 61, 62 Recent advances in DL have shown strong potential in leveraging H&E-stained slides to predict biomarkers evaluated routinely in IHC, offering a possibly more consistent and faster alternative, potentially reducing turnaround time up to 45%.63, 64, 65, 66, 67, 68, 69

When considering ER prediction from H&E-stained WSIs, a study including ∼460 patients in the external test demonstrated AUC values up to 0.90.⁶³ Another study demonstrated an accuracy of almost 85% in predicting ER status from H&E-stained WSI on an independent set of nearly 290 cases.⁶⁴ When evaluating PR receptor, in a multicohort validation setting, PR status prediction with AI tools reached concordance levels of 86%-91% with manual scoring, while HER2 negativity was correctly identified in 91%-100% of cases.⁶⁵

A multi-institutional study evaluated a DL system using WSIs from 7950 breast cancer patients.⁶⁶ The model identified 30% of cases as HR-positive, achieving a specificity and PPV of almost 100%, demonstrating its effectiveness in determining HR status without IHC. Moreover, through re-evaluation of cases highlighted as potential false negatives, ∼30 misclassified cases were correctly identified.⁶⁶

In parallel, a few studies also evaluated the capability of DL to predict PD-L1 expression from H&E slides.^67,68 These studies reported strong predictive performance for PD-L1 expression, with AUC values exceeding 0.90 in a large training cohort of ∼3300 patients, with superimposable results in two different validation cohorts.⁶⁷

These algorithms, capable of inferring IHC results directly from standard H&E slides, could provide a more affordable alternative in under-resourced settings. In such regions, local laboratories may rely on low-cost, single-slide scanners rather than expensive high-throughput whole-slide scanners. The digitized slide could then be shared to a better-equipped reference centre, where the computational analysis is carried out. In high-income countries, however, where IHC is routinely carried out, the added value of such tools may be more limited.

Prediction of molecular alteration and gene expression from H&E-stained WSIs

BRCA and HRD status

Breast cancer susceptibility genes 1 and 2 (BRCA1 and BRCA2) are key tumour suppressor genes involved in DNA repair via homologous recombination, and their germline mutations are associated with an increased risk of breast cancer, with a cumulative risk to the age of 80 years being almost 70%.⁷⁰ Identifying BRCA1/2 mutations is not only critical for guiding personalized therapies, but also for implementing preventive strategies in high-risk patients and providing oncogenetic counselling.⁷¹

However, standard genetic testing remains expensive and inaccessible in resource-limited settings, with the risk of delaying therapeutic decisions exacerbating health care inequalities. Due to this, recent research explored H&E-based AI approaches to infer BRCA1/2 mutation status and homologous recombination deficiency (HRD), offering a faster and cheaper alternative.72, 73, 74

One study employed a DL model to predict germline BRCA1/2 mutations from WSIs. The model achieved AUC values of up to 0.77 at higher magnifications (×20 and ×40) in the validation set.⁷² The tumour grade influences the algorithm’s performance, with improved accuracy observed in grade 3 across different magnifications, suggesting that image resolution captures complementary histological features.

Another study introduced an attention-based model, which showed improved accuracy, achieving an AUC of 0.82 in both validation sets.⁷³ Importantly, this model linked morphological features such as tumour-infiltrating lymphocytes and nuclear characteristics to BRCA1/2 mutation status.⁷³

A third investigation focused on predicting HRD, a key biomarker for response to platinum chemotherapy and poly(ADP-ribose) polymerase inhibitors.^74,75 This model reached an AUC of 0.86 in the internal dataset after correction for bias, and identified histological features like laminated fibrosis and clear-cell morphology associated with HRD-positive tumours.⁷⁴

Genomic signatures

The Oncotype DX Breast Recurrence Score (ODX) is the most adopted genomic test in early-stage HR-positive/HER2-negative breast cancer, being validated in prospective randomized clinical trials (RCTs).^76,77 It analyses 21 tumour-related genes to produce a recurrence score from 0 to 100, directly linked to recurrence risk and it is included in international guidelines.⁷⁸ Scores between 0 and 25 indicate low risk of recurrence, while scores >25 imply high risk, often leading to chemotherapy recommendation. Unfortunately, its high price and slow availability limited its wide use, and any effort to predict it by clinicopathological factors have not shown sufficient efficacy.⁷⁹ Due to this, studies have also been conducted in order to predict the ODX value based on WSIs.80, 81, 82

One DL model utilizes a CNN for predicting ODX from WSI feature extraction and spatial aggregation, integrated with clinicopathological data.⁸⁰ This method demonstrated great accuracy in classifying patient risk categories based on the ODX, achieving an AUC of 0.85 in an external test set of ∼400 patients.⁸⁰

Similarly, Orpheus, a bimodal DL tool developed by StratifAI® on >6000 cases evaluating WSIs and pathological reports, outperformed traditional clinicopathological nomograms by identifying high-risk as per ODX, with an AUC of 0.89.⁸¹ Interestingly, the tool was more accurate than ODX in predicting metastatic recurrence score in patients with low-risk ODX.⁸¹

Moreover, another DL model trained on a dataset of ∼150 breast cancer patients, reached an AUC of 0.86 in predicting ODX risk categories from histopathology alone.⁸²

Prognostic data

Estimating recurrence risk in early-stage breast cancer (eBC) is essential for personalized treatment. Traditional pathological prognostic factors (e.g. Ki-67, tumour grade) are insufficiently reliable due to tumour heterogeneity and subjective interpretation by pathologists. CP enables scalable prognostic assessments from H&E-stained WSIs extracting pattern beyond pathologists’ capability.

Multiple companies developed prognostic tools, especially for HR-positive/HER2-negative eBC (see Table 2). Even though few tools reached a clinical validation level of IB (level 1B evidence, i.e. external validation with archived specimen from at least one RCT, such as Spotlight Medical®) and some are used in clinical routine (e.g. Stratipath® in Karolinska Institute), most of these tools require prospective validation in proper RCTs to prove their clinical utility.

Table 2.

Non-exhaustive list of commercial companies developing models for prediction of prognosis

Company	Product	Clinical endpoint	Intended use	CE-IVD marked	Sources
Histofy®	BRACE marker	MFS DSS	Guide adjuvant CT		83
Owkin®	RlapsRisk BC	MFS	Guide adjuvant CT	✓	85
PreciseDx®	Precise Breast	Recurrence	Guide adjuvant therapy		87
Stratipath®	Stratipath Breast	Histological grade PFS	Guide adjuvant CT	✓	84
Spotlight Medical®	NA	DRFI IDFS OS	Guide adjuvant CDK4/6 inhibitor		86
StratifAI®	Breast CDx	ODX RS	Guide adjuvant CT		81

Of note, none of these tools had been validated in a prospective randomized clinical trial so far.

CDK4/6, cyclin-dependent kinase 4 and 6; CE-IVD, Conformité Européenne – in vitro diagnostic marked; CT, chemotherapy; DRFI, distant recurrence-free interval; DSS, disease-specific survival; IDFS, invasive disease-free survival; MFS, metastasis-free survival; NA, not applicable; ODX RS, Oncotype DX recurrence score; OS, overall survival; PFS, progression-free survival.

A compelling illustration is the BRACE marker, which evaluates tumour heterogeneity using DL in HR-positive/HER2-negative eBC.⁸³ In two cohorts, the largest of ∼2100 patients, BRACE accurately distinguished high- from low-risk patients in terms of distant metastasis-free survival and breast cancer-specific survival, with a concordance-index (C-index) of 0.73 and 0.84, respectively.⁸³

In a retrospective study involving >2700 patients, Stratipath Breast®, an AI-based tool of H&E WSIs, was evaluated to measure its prognostic performance.⁸⁴ The tool divided patients into high- and low-risk groups and demonstrated an independent prognostic value, especially in the subgroups represented by HR-positive/HER2-negative patients and grade 2 disease. In fact, high-risk patients for Stratipath® Breast® showed significantly worse outcomes in these two subgroups as compared with low-risk patients when considering PFS (hazard ratio 2.76, P < 0.001 and hazard ratio 2.20, P = 0.009, respectively).

RlapsRisk BC is an AI-based tool designed to predict 5-year metastasis-free survival (MFS) from H&E WSIs in patients with HR-positive/HER2-negative eBC divided into low and high risk based on a 5% MFS probability threshold without requiring local annotations by a pathologist.⁸⁵ Validated on ∼1170 patients included in the CANTO trial, a French multicentric national cohort, the model outperformed conventional clinicopathological indicators, with a C-index of 0.81 and 0.76, respectively. Moreover, the tools added ∼5% in both sensitivity and specificity as compared with clinicopathological items alone.⁸⁵

Bidard et al. recently developed a predictive model externally validated in >600 high-risk HR-positive/HER2-negative eBC.⁸⁶ This approach combined histological features and four clinical variables, dividing patients into low risk and not low risk of recurrence. External validations on archived data from two prospective clinical trials showed that this risk stratification significantly surpassed clinical-only models in predicting distant recurrence-free survival, invasive disease-free survival and overall survival (OS) (hazard ratio for OS 0.35, 95% CI 0.13-0.97, P = 0.044).⁸⁶ The authors proposed a non-escalation therapeutic strategy for the identified low-risk group of patients.

A retrospective single-centre study evaluated the activity of an AI tool in prediction of prognosis from WSIs of ∼520 eBC in the validation cohort.⁸⁷ The AI assay demonstrated a C-index of 0.75 in the validation cohort, surpassing the performance of clinical models. When considering recurrence prediction at 6 years, the hazard ratio was 4.4 in the validation cohort between low and high risk (95% CI 2.7-7.1, P < 0.001).

Discussion

CP can simplify and enhance pathologists’ work, potentially opening a wide range of opportunities for diagnostic accuracy and enabling more personalized treatment approaches, though several barriers to its adoption persist.

A major challenge is demonstrating clinical utility for AI-derived biomarkers: while prospective RCTs remain the gold standard, their implementation is limited because they are expensive and time-consuming.⁸⁸ Ethical concerns also persist, because using an unvalidated biomarker in an RCT to skip standard treatment may expose patients to inadequate care.⁸⁸ This problem can be mitigated by using archived specimens from well-structured RCTs for external validation.⁸⁸

In addition, digital transformation of the pathology department comes with important economic investments. Although cost savings can be achieved in the medium/long term, the upfront expenditure represents a barrier for small or mid-size institutions.⁸⁹

Moreover, AI models in oncology may show biases, potentially limiting accuracy and equity in health care.⁹⁰ Uncovering and mitigating biases is crucial to create reliable AI tools, using comprehensive evaluations with diverse frameworks and challenging datasets to reveal hidden biases.⁹⁰

Another limitation is the lack of business models for many AI-based diagnostic applications, thus limiting the adoption of these techniques and contrasting with areas like companion diagnostics, where reimbursement pathways are clearer.⁹¹

To address these challenges, the STARD-AI guidelines provide a comprehensive framework to improve the reporting quality and transparency of diagnostic accuracy studies involving AI.⁹²

Lastly, an important limitation of DL is the lack of interpretability of the results. Even though interpretable AI models do exist, most of the current models evaluating tasks such as predicting treatment response or patient prognosis cannot be directly verified by a pathologist or confirmed through alternative tests like IHC or genomic assays.^86,93 This ‘black box’ nature raises transparency and regulatory concerns. Efforts to develop interpretable AI aim to provide clearer insights without compromising predictive performance.⁷

Recent developments in agent-based AI systems supporting oncologists during multidisciplinary tumour board evocate earlier efforts like IBM Watson, which struggled to reach clinical expectations due to limited real-world integration.⁹⁴ Today, however, AI has evolved into a more adaptive and interactive aid. New frameworks combining large language models with guideline consultation showed encouraging results. Recent studies including a recent work by Ferber et al., suggest these tools can match closely with oncologists’ decision.⁹⁵ We anticipate a wave of proof-of-concept studies that will help move this technology closer to meaningful clinical deployment.

In conclusion, as technology evolves and evidence accumulates, CP is poised to become an integral part of oncological routine, even though for now clinical validation is not demonstrated so far.