Enhancing lymph node metastases assessment in breast cancer post-neoadjuvant therapy using artificial intelligence-driven diagnostics

Abstract

Background

Neoadjuvant therapy (NAT) is crucial for locally advanced breast cancer, but post-NAT lymph node assessment is challenging due to histological changes. Current methods like immunohistochemistry (IHC) are labor-intensive and imprecise in distinguishing isolated tumor cells (ITCs), micro-metastases (Micro), and macro-metastases (Macro). We aimed to develop and validate an AI-driven model for precise classification of lymph node metastasis status (negative, ITC, Micro, Macro) in breast cancer patients post-NAT.

Methods

We used a weakly supervised Clustering-constrained Attention Multiple Instance Learning (CLAM) framework to analyze 7764 lymph node samples from 7 cohorts. The CLAM model identifies high-diagnostic-value subregions within whole-slide images (WSIs) and generates high-resolution interpretability heatmaps. Performance was evaluated using binary and multi-class metrics, with external validation on diverse datasets. A human-AI comparative analysis was conducted on 24 patient-derived lymph node sections.

Results

The AI model achieved an AUROC of 0.97 (95% CI: 0.962–0.977) in binary classification and an overall accuracy of 0.8436 (95% CI: 0.8282–0.8562) for multi-class differentiation. In the human-AI comparison, the model outperformed junior pathologists, reducing diagnostic discrepancies by 83%.

Conclusion

This study establishes a robust AI model that significantly improves the accuracy and efficiency of post-NAT lymph node metastasis assessment in breast cancer, automating classification and reducing pathologist workload.

Introduction

Breast cancer is the most common malignancy among women worldwide, with neoadjuvant therapy (NAT) increasingly used to reduce tumor size and improve surgical outcomes in locally advanced case [1, 2]. The pathological status of the lymph nodes following NAT is a critical factor influencing patients’ pathological complete response (pCR) and long-term prognosis [3]. In particular, the presence of sentinel lymph node (SLN) metastasis directly determines subsequent surgical approaches and treatment strategies [4]. Therefore, accurate assessment of lymph node status post-NAT is of significant clinical value.

Evaluating lymph node metastasis following NAT presents numerous challenges. Based on the size of metastatic deposits, lymph node involvement is typically categorized into 4 groups: isolated tumor cells (ITCs), micro-metastases (Micro), macro-metastases (Macro), and no metastasis [5, 6]. According to the latest standards for neoadjuvant breast cancer assessment, as outlined in STEEP 2.0, a pathological complete response is defined as the complete absence of residual tumor cells in both breast tissue and lymph nodes [7]. However, NAT can induce various changes within the lymph nodes, such as tumor cell regression, stromal fibrosis, collagenization, and histiocytic proliferation. These alterations may interfere with the pathologist’s visual assessment, potentially compromising the accuracy and efficiency of diagnosis [8].

Currently, most methods for assessing lymph node metastasis rely on pathologist-based visual assessment, which is often limited to distinguishing between the presence or absence of metastasis, with insufficient accuracy in differentiating ITCs, Micro, and Macro [9]. Moreover, there is a lack of comprehensive data on metastasis evaluation following NAT [10, 11]. Traditional methods are also constrained by limited data volume and a lack of robust validation across diverse clinical settings, which impacts their precision and reliability in clinical applications. Therefore, the development of a method capable of accurately distinguishing different types of metastasis and improving diagnostic efficiency is imperative.

Artificial intelligence (AI), particularly deep learning techniques, has made groundbreaking advances in the field of pathology in recent years. AI can analyze large volumes of pathological images to automatically detect and classify tumor cells, significantly improving the diagnostic accuracy of lymph node metastasis in breast cancer [4, 12, 13]. Existing deep learning algorithms, such as convolutional neural networks (CNNs), have been applied to the analysis of breast cancer histopathological slides, enabling the automatic identification of metastatic cancer cells and providing diagnostic assistance [14]. Building on this, our study developed a deep learning model specifically designed for evaluating lymph node metastasis following NAT, intending to enhance both the accuracy and efficiency of pathological diagnosis.

This study employed deep learning techniques to automatically classify lymph node metastasis types post-NAT, with a particular emphasis on accurately distinguishing ITCs, Micro, and Macro. A large-scale dataset was meticulously curated for model training and validation, and its robustness and broad applicability were further demonstrated through external test using multi-center data. This research introduces an innovative intelligent tool for the clinic-pathological evaluation of breast cancer following NAT, offering significant clinical implications, such as improved diagnostic accuracy and personalized treatment planning.

Methods

Data acquisition and dataset description

This study involved the collection of digital WSIs of lymph nodes from 648 breast cancer patients who underwent NAT. Of these, 310 cases were sourced from the Fourth Hospital of Hebei Medical University (H4), while the remaining 338 cases were obtained from Affiliated Hospital of Hebei University (BD), Xingtai People’s Hospital (XT), First Hospital of Qinhuangdao (QHD), Cangzhou Integrated Traditional Chinese and Western Medicine Hospital (CZ), and Handan First Hospital (HD). Patients were excluded according to the following criteria: (1) incomplete clinical or pathological information; (2) poor-quality WSIs (e.g., severe blurring, artifacts, or incomplete tissue); (3) patients who received less than 3 cycles of NAT. After standardized enrollment and exclusion screening, a total of 447 eligible patients with 7264 lymph node specimens were eventually incorporated into the final analysis. The selection process for patients enrolled in this study is illustrated in Appendix 1. These samples were meticulously diagnosed by experienced pathologists according to the 8th edition of the AJCC staging system, classifying them into categories of negative, ITCs, Micro, and Macro. Additionally, 500 lymph node samples from the public available CAMELYON17 dataset (https://camelyon17.grand-challenge.org/Data/), which also involved post-neoadjuvant chemotherapy (NAC) specimens, were incorporated. All samples were stratified into 7 cohorts based on their source: H4, XT, QHD, CZ, HD, BD, and CAMELYON17 (C17). These samples were allocated into a training dataset (dataset 1 from C17, H4 and HD), a validation set (dataset 2 from H4), an internal test dataset (dataset 3 from H4 and HD), and an external test (dataset 4 from BD, QHD, XT, and CZ). A separate dataset (dataset 5 from H4) was reserved for the comparative analysis between the AI model and pathologists. This structured allocation was designed to facilitate the development of an AI model for lymph node recognition.

Cytokeratin immunohistochemistry

For false positive cases, we performed immunohistochemical staining for validation. The primary antibody CK-pan (Clone: AE1/AE3) were purchased from ZSGB-BIO, China, as a working solution. Immunohistochemistry was performed using Ventana Benchmark XT. In our study, a false positive case was strictly defined as a lymph node that met all three of the following criteria: (1) Classified as positive (any of ITC, micro, or macro) by the AI model; (2) Diagnosed as negative by the consensus of two senior pathologists on H&E-stained slides; (3) Confirmed as negative by CK-pan immunohistochemistry (IHC), with no detectable tumor cells.

Image pre-processing

We first performed tissue detection to remove the large areas of white background for each WSI. The tissue detection was based on binary thresholding, median blurring, and morphological operations. We extracted 256 × 256 patches without overlapping from each slide. A ResNet-50 [15] is used to extract the patch features from each slide, resulting in a N× 1024 matrix, where N is the number of patches extracted from a slide.

Multiple instance learning

Multiple instance learning (MIL) [16] is a popular weakly supervised learning algorithm for WSI recognition. The goal of MIL is to perform classification on data that is arranged in bags of instances (patches). Each instance is either positive or negative, these instance-level labels are not available during training, but only bag-level (slide-level) labels are available [16–18]. A bag (slide) is labeled as positive if any of the instances in it are positive, and negative otherwise.

CLAM

Clustering-constrained-attention multiple-instance learning (CLAM) [16] is a deep-learning-based weakly supervised method that uses attention-based learning to identify sub-regions of high diagnostic value to accurately classify whole slides and instance-level clustering over the identified representative regions to constrain and refine the feature space. It can be used to localize well-known morphological features on WSIs without the need for spatial labels. For whole-slide-level learning without annotation, CLAM uses an attention-based pooling function to aggregate patch-level features into slide-level representations for classification. At a high level, during both training and inference, the model examines and ranks all patches in the tissue regions of a WSI, assigning an attention score to each patch, which informs its contribution to the collective slide-level representation for a specific class.

Specifically, for a WSI represented as a bag of K patches, the instance-level embedding corresponding to the k-th patch is denoted as . The first fully connected layer in CLAM further compresses each fixed patch-level representation to vector . The attention network consists of several fully connected layers, consider the first two layers of the attention network and and collectively as part of the attention backbone shared by all classes, the attention network then splits into N parallel branches . Accordingly, the attention score of the k-th patch for the i-th class is given by

The slide-level representation aggregated per the attention score distribution for the i-th class is given as follows:

Then, the slide-level score is given via the classifier layer by . The model was trained with a learning rate of 2e-4 with a decay of 1e-5, and a dropout was performed after each layer of the attention backbone for regularization.

Statistical analyses

This study evaluates the performance of machine learning binary classification models using multiple metrics, including AUC (Area Under the Curve), Sensitivity, Specificity, PPV (Positive Predictive Value), NPV (Negative Predictive Value), F1 score, and Balanced Accuracy. AUC represents the area under the ROC (Receiver Operating Characteristic) curve and measures the overall performance of a classification model across various thresholds. Its values range from 0.5 to 1, with higher values indicating stronger discrimination between positive and negative samples. Sensitivity reflects the proportion of actual positive samples correctly identified as positive, while Specificity evaluates the proportion of actual negative samples correctly identified as negative. PPV assesses the probability that samples predicted as positive are truly positive, whereas NPV evaluates the probability that samples predicted as negative are truly negative. The F1 score is the harmonic mean of Precision and Recall (equivalent to Sensitivity), providing a comprehensive evaluation of the model’s accuracy and completeness in predicting positive samples. Balanced Accuracy, calculated as the average of Sensitivity and Specificity, mitigates the bias caused by imbalanced class distributions and offers a more holistic reflection of classification performance.

For multi-class classification tasks, the diagnostic performance for a specific class is assessed by designating that class as the positive outcome and merging the remaining three classes as the negative outcome. This approach enables an evaluation of the classification performance for individual categories.

Results

Patients and clinical‐pathological characteristics

The 6 cohorts were derived from 447 excised breast cancer specimens collected from patients who had undergone NAC at 6 different medical centers. These specimens resulted in a total 7264 lymph nodes. The clinicopathological characteristics of the enrolled patients are summarized in Table 1.

Table 1.

Baseline demographic and clinical characteristics of patients

Variable	n = 447	%
Age
< 60	322	72.0%
≥ 60	123	27.5%
Unknown	2	0.4%
Histologic type
NST	398	89.0%
Special type	17	3.8%
Mixed	27	6.0%
Unknown	5	1.1%
Nottingham grade
I	10	2.2%
II	373	83.4%
III	53	11.9%
Unknown	11	2.5%
Immunophenotype
Luminal A	49	11.0%
Luminal B	175	39.1%
HER2 Positive	95	21.3%
Triple Negative	91	20.4%
Unknown	37	8.3%
RCB grade
0	81	18.1%
I	48	10.7%
II	163	36.5%
III	153	34.2%
Unknown	2	0.4%
Lymph node metastases
Yes	306	68.5%
No	141	31.5%
Neoadjuvant
Yes	447	100.0%
No	0	0.0%

Abbreviations: NST, invasive carcinoma of no special type; RCB, residual cancer burden

Proposed AI model and dataset characteristics

As previously described, we collected 7767 axillary lymph node specimens from 6 medical centers and one public database to develop a multi-class AI model for breast cancer patients following NAC. The model’s performance was rigorously validated through human–machine comparison (Fig. 1A). To preprocess the WSI data, we employed ResNet-50 for feature extraction. Given the absence of instance-level annotations, we adapted the MIL framework, organizing patches into bags where a bag was labeled “positive” if any constituent patch contained malignancy aligning with slide-level annotations (Fig. 1B). Subsequently, we integrated the CLAM method, which leverages attention mechanisms to prioritize diagnostically relevant regions and aggregates patch-level features into a robust slide-level representation. Furthermore, CLAM incorporates instance-level clustering constraints to preserve morphological fidelity, enabling accurate identification of key pathological features without spatial annotation (Fig. 1C).

Fig. 1.

Workflow of the AI-driven lymph node classification. A Overview of the AI-assisted diagnostic pipeline. B Lymph node regions are segmented via binary thresholding (left) and tiled into patches for analysis (right). C Patch features are extracted once using a pre-trained ResNet-50 and aggregated by a CLAM attention network to generate node-level predictions. D Summary of data usage

From the 7764 lymph node samples across 7 multi-center cohorts, we divided the data into training, validation, and internal test datasets in approximately a 4:1:2 ratio. Additionally, 999 lymph node samples, originating from 4 distinct medical centers within these cohorts, were selected as the external test dataset. The cohort origins and lymph node metastasis statuses (negative, ITCs, Micro, Macro) for each dataset are detailed in Fig. 1D.

Performance of AI model

We first assessed the AI model’s ability to distinguish between metastatic and non-metastatic lymph nodes using binary classification. In internal validation, the model achieved consistently high performance across cohorts H4 and HD, with AUCs of 0.977 (95% CI: 0.967–0.986) and 0.962 (95% CI: 0.937–0.987), respectively. External validation across four independent multicenter cohorts further confirmed its generalizability, yielding AUCs of 0.97 (CZ), 0.977 (XT), 0.92 (QHD), and 0.967 (BD) (Fig. 2A–G, I)—demonstrating robust diagnostic accuracy despite inter-institutional heterogeneity. Overall, on the combined external test set, the model attained an AUC of 0.97 (95% CI: 0.962–0.977) (Fig. 2H), confirming its strong and consistent performance in identifying lymph node metastasis.

Fig. 2.

Performance of the AI model in binary classification of lymph node metastasis. A Confusion matrix for the overall test dataset. B Confusion matrix for the internal test dataset H4 cohort. C Confusion matrix for the HD cohort. D–G Confusion matrices for the external test datasets: CZ, XT, QHD, and BD cohorts, respectively. H AUC curve illustrating the AI model’s performance in predicting lymph node metastasis across all test datasets from different medical centers. I Radar chart summarizing the multi-metric performance of the AI model across diverse test cohorts

In the four-class classification (negative, ITC, Micro, Macro), the model achieved an overall accuracy of 0.8436 (95% CI: 0.8282–0.8562), with a Kappa value of 0.7271, demonstrating strong concordance with the gold-standard pathological assessments. The classification accuracy for negative nodes was 92.47%, for ITC was 84.64%, for Micro was 83.92%, and for Macro was 89% (Fig. 3). Notably, the model exhibited exceptional precision in detecting Macro metastases, aligning with clinical priorities to assess larger metastatic foci. Furthermore, the model also performed admirably in identifying micro and ITC, which are often more challenging to diagnose clinically.

Fig. 3.

Performance of the AI model in 4 classifications of lymph node metastasis. A Confusion matrix for the overall test dataset. B Confusion matrix for the internal test dataset H4 cohort. C Confusion matrix for the HD cohort. D–G Confusion matrices for the external test datasets: CZ, XT, QHD, and BD cohorts, respectively. H Heatmap summarizing the AI model’s multi-center performance across key diagnostic metrics

Comparison of AI model and pathologists in the assessment of lymph node metastasis status

24 patient-derived lymph node sections from the H4 cohort were selected for human-AI comparative analysis to validate the diagnostic accuracy of the AI model and test results for each patient are depicted in Fig. 5. Among the 70 positive lymph nodes identified by the gold-standard, the AI model’s assessment differed for only 1 lymph node. In contrast, a senior pathologist had discrepancies in 14 lymph nodes, while 3 mediate-level pathologists each had 13 discrepancies. The 3 junior pathologists reported inconsistencies in 24, 13, and 15 lymph nodes, respectively (Fig. 4A). Furthermore, further analysis of the discrepancies in the classification of positive lymph node categories indicates that, relative to the gold-standard, both the AI model and pathologists exhibited the highest number of misclassifications in the ITCs category, followed by Micro (Fig. 4B, C). In the case of Macro, the AI model’s assessments were in complete agreement with the gold-standard. However, senior pathologists recorded 10 discrepancies, while each of the 3 intermediate and 3 junior pathologists noted 9 discrepancies (Fig. 4D). A detailed analysis revealed that the majority of errors made by the pathologists involved misclassifying Macro as Micro-metastases. The specific performance of the AI model and pathologists across the 3 categories of positive lymph nodes is illustrated in appendix 2.

Fig. 5.

Heatmap depicting the AI model’s detection of ITCs. A, D Lymph node slide with ITCs (scale bar = 2 mm). B, E Enlarged view of the localized area within the lymph node containing ITCs (scale bar = 2 μm). C, F Heatmap analysis of the localized area within the lymph node containing ITCs

Fig. 4.

Comparison of discrepancies in detecting positive lymph node metastasis between the AI model and pathologists. A Positive cases heatmap. Each column represents the lymph node data for an individual patient. The numbers within the heatmap indicate the total count of positive lymph nodes among all samples for each patient, encompassing ITCs, Micro, and Macro. B–D Comparison of inconsistencies in ITCs, Micro and Macro case diagnoses between the AI model and pathologists, against the gold-standard

Attention map analysis of the AI classification

In our study, we observed notable chemotherapeutic responses in lymph nodes, including necrosis, pronounced fibrotic/collagenous changes, histiocytic hyperplasia, and reactive lymphocyte infiltration. These morphological alterations significantly compromise the accuracy and interobserver consistency of pathologists’ visual assessments. Furthermore, through systematic analysis and comparative evaluation of challenging cases, complemented by attention-based heatmaps that visually highlight the AI model’s focal regions of diagnostic emphasis—we demonstrated robust performance of our algorithm in these complex lymphatic specimens (Figs. 5, 6 and 7). As illustrated in these figures, regions with heightened attention weights (represented by intensified red hues) demonstrated spatial co-localization with ITCs, Macro, and Micro confirmed via histopathological magnification, whereas normal lymph node tissue regions were annotated in green hues.

Fig. 6.

Heatmap Depicting the AI Model’s Detection of Macro. A, D Lymph node slide with Macro (scale bar = 2 mm). B, E Heatmap analysis of the localized area within the lymph node containing Macro (scale bar = 2 mm). C, F Enlarged view of the localized area within the lymph node containing Macro (red areas) and normal lymph node tissue (green areas) (scale bar = 100 μm)

Fig. 7.

Heatmap Depicting the AI Model’s Detection of Micro. A, D Lymph node slide with Micro (scale bar = 2 mm). B, E Heatmap analysis of the localized area within the lymph node containing Micro (scale bar = 2 mm). C, F Enlarged view of the localized area within the lymph node containing Micro (red areas) and normal lymph node tissue (green areas)

Discussion

NAT has become the primary treatment strategy for patients with locally advanced breast cancer, due to its effectiveness in reducing tumor burden preoperatively [1, 19]. This approach enhances rates of breast conservation and reduces the extent of axillary surgery. The RCB system, strongly endorsed by the Breast International Group-North American Breast Cancer Group (BIG-NABCG), is recommended for the assessment and reporting of tumor staging in breast tissue specimens following NAC [20]. Within this framework, the quantity of lymph node metastases and their maximum diameter are critical components. However, Micro and ITC detected post-NAT have been identified as predictors of poorer survival outcomes [10, 21, 22]. Therefore, even a minimal presence of cancer cells in lymph nodes post-therapy should not be interpreted as a pCR [3, 23]. According to the WHO’s 2012 classification of breast tumors, ITC in lymph nodes after NAC should be considered lymph node-positive [24]. Consequently, all excised lymph nodes require meticulous pathological examination. However, this manual histopathological analysis is labor-intensive, repetitive, and potentially prone to inaccuracies, as pathologists may overlook minute metastases [24, 25]. Moreover, distinguishing between ITCs and Micro poses significant challenges due to low inter-observer consistency [26]. Nonetheless, precise differentiation of lymph node metastasis types is crucial, as it significantly influences subsequent therapeutic decisions and clinical outcomes for patients [27, 28].

This study proposes a solution to address lymph node metastasis post-NAT by employing a deep learning-based weakly supervised method known as CLAM, to develop an AI model for classifying lymph node metastatic status using H&E stained sections. This model not only efficiently determines whether lymph nodes are infiltrated by tumor cells performing a binary classification of metastatic presence but also accurately categorizes the type of metastasis (i.e., Macro, Micro, and ITCs), thereby significantly enhancing the accuracy and efficiency of lymph node metastasis evaluation. By automating this process, the model reduces the workload of pathologists in lymph node examination, achieving diagnostic performance that is comparable to, if not surpasses, that of human pathologists.

AI-based on digital pathology has been extensively explored for evaluating lymph node metastases in breast cancer [12, 29–31]. The CAMELYON16 grand challenge, held in 2015, aimed to challenge evaluated AI algorithms for the automated detection of metastases in H&E-stained WSIs of lymph node sections. 7 of 32 evaluated AI algorithms showed greater discrimination than a panel of 11 pathologists who operated in a simulated, time-constrained environment. In this challenge, the best-performing pathologist in the study was outperformed by the best performing algorithm on the study set [32]. However, the sensitivity rates for detecting ITCs during the CAMELYON challenge were reported to be below 40%, highlighting the difficulty of accurately diagnosing small lymph node metastasis types, such as Micro and ITCs, which can lead to mis-assessment. Liu et al. developed an algorithm named LYmph Node Assistant (LYNA) using 270 slides from the CAMELYON16 challenge dataset to detect macro- and Micro in SLN biopsy WSIs stained with H&E [29]. LYNA achieved a slide-level AUC of 99% and a tumor-level sensitivity of 91% at one false positive per patient on the 129 slides of the CAMELYON16 evaluation dataset. The sensitivity for detecting Macro- and Micro was 71.6%, comparable to the 72% sensitivity achieved by 11 pathologists. However, most lymph nodes in these studies were from patients who had not undergone NAT. Post-NAC, lymph nodes often undergo various reactions, making metastatic nodes more challenging to identify.

In our study, we utilized the largest dataset of lymph nodes post-NAT to date, comprising 7764 lymph nodes, to train a high-performance classification and recognition model. For the binary classification of breast cancer cell metastasis, the AI model achieved an AUC of 0.97 (95% CI: 0.962–0.977) for the overall test dataset, with a sensitivity of 0.98, specificity of 0.87, F1 score of 0.89, and accuracy of 0.92. In a more detailed 4-class classification, the model achieved an accuracy of 0.8436 (95% CI: 0.8282–0.8562) for the overall test dataset and a Kappa value of 0.7271, indicating substantial agreement with the gold-standard diagnoses. The sensitivity for subcategories of lymph node metastases was 0.7538 for ITCs, 0.7421 for Micro, and 0.8257 for Macro. This performance is similar to the Paige breast lymph node (Paige BLN) model, developed by Retamero et al. [33], which was trained on over 32,000 WSIs from more than 8,000 patients. Paige BLN demonstrated an overall sensitivity of 92.8% for lymph node metastasis detection, with an independent sensitivity of 78% for ITCs, specificity of 94.9%, PPV of 92.8%, and NPV of 94.9%. However, the Paige BLN study also had a relatively small proportion of lymph nodes from patients who had undergone NAC.

Additionally, Li et al. [34] developed a “metastasis AI” detection app to screen lymph nodes for metastases in breast cancer. In their evaluation of 81 positive lymph nodes (including 40 Macro, 36 Micro, and 5 ITCs), the app achieved an overall sensitivity of 100%, specificity of 78.5%, PPV of 68.1%, and NPV of 100%. This performance is similar with our AI model, which also achieved an overall sensitivity of 100% and specificity of 78.5% on the test dataset. However, Li et al.’s study also had a limited number of NAC samples, but their app has the advantage of finely segmenting tumor-invaded lymph node regions, whereas our model can only identify and classify lymph node metastases at the slide level. Considering the complex changes in lymph nodes post-NAC, such as necrosis and pronounced fibrosis and collagenization, these alterations can lead to misinterpretation of ITCs or Micro by pathologists. However, in this study, the AI model demonstrated excellent performance in accurately identifying these phenomena. The visualization through heatmaps also illustrated the AI model’s basis for judgment (Figs. 5, 6 and 7). Moving forward, our research team will continue to optimize the AI model, such as by training it for precise segmentation of tumor invasion within lymph nodes, enabling pathologists to focus on specific areas of interest and reducing the likelihood of oversight in the context of NAC. Furthermore, this study demonstrates that AI significantly enhances the diagnostic accuracy of lymph node metastases compared to manual evaluation, particularly in assisting junior pathologists. This study also has some limitations. Firstly, it relies heavily on pathologist expertise for data annotation. As mentioned, our AI model requires further refinement for precise segmentation of tumor invasion areas within lymph nodes. Additionally, assessing how the AI model enhances pathologist efficiency in diagnosing lymph node metastases is crucial. Developing a real-time, interactive AI tool for pathologists could revolutionize pathology workflows. Such a tool could provide immediate feedback during slide examination, highlighting suspicious areas and suggesting classifications for lymph node metastases. This would enhance diagnostic accuracy and reduce slide review time, especially in high-volume settings. Integrating AI into digital pathology platforms could also facilitate remote diagnostics, allowing pathologist collaboration across institutions.

Conclusion

In summary, our research presents the first AI classification and recognition model entirely focused on lymph node metastases in breast cancer post-NAT, achieving performance in various categories that rivals existing AI models. Overall, we have made advancements in the study of breast cancer NAT lymph node metastases, offering pathologists an efficient and accurate AI tool for assessing the therapeutic efficacy of breast cancer NAT. Future work will focus on refining the model’s capabilities, integrating multi-modal data, and validating its performance in prospective clinical trials to ensure its clinical utility and impact on patient care.

Abbreviations

AI

Artificial intelligence

NAT

Neoadjuvant therapy

IHC

Immunohistochemistry

ITCs

Isolated tumor cells

Micro

Micro-metastases

Macro

Macro-metastases

CLAM

Clustering-constrained attention multiple instance learning

WSIs

Whole-slide images

pCR

Pathological complete response

SLN

Sentinel lymph node

CNNs

Convolutional neural networks

NAC

Neoadjuvant chemotherapy

MIL

Multiple instance learning

AUC

Area under the curve

PPV

Positive predictive value

NPV

Negative predictive value

ROC

Receiver operating characteristic

CI

Confidence interval

Appendix 1

See Fig. 8.

Fig. 8.

Patients selection process. This study involved the collection of WSI of lymph nodes from 648 breast cancer patients who underwent NAT. Of these, 310 cases were sourced from the Fourth Hospital of Hebei Medical University (H4), while the remaining 338 cases from five additional institutions: Affiliated Hospital of Hebei University (BD), Xingtai People's Hospital (XT), First Hospital of Qinhuangdao (QHD), Cangzhou Integrated Traditional Chinese and Western Medicine Hospital (CZ), and Handan First Hospital (HD). Following a rigorous selection process, 447 patients were ultimately included, resulting in a total of 7264 lymph node samples. These samples were systematically stratified into distinct datasets to facilitate AI model construction, including training, validation, and testing stages

Appendix 2

see Fig. 9.

Fig. 9.

Heatmaps of positive lymph node metastasis identified by AI model and pathologists. Each column represents the lymph node data for an individual patient. The numbers within the heatmap indicate the total number of positive lymph nodes for each patient. Panels A–C display heatmaps of positive ITCs, Micro, and Macro cases, respectively

Author contributions

YD and JY were responsible for data curation, writing and editing the manuscript. ML, LK and LH conducted data analysis and figure integration. JL, YW and XX performed the histopathologic analysis. XL built the models of the AI. MZ, PW, SL, YW and XX provided the lymph nodes slides. ZL and YL contributed to the conceptualization, supervision, and reviewing the manuscript. All the authors read and approved the final manuscript. All authors had full access to the data and had final responsibility for the decision to submit for publication. ML, LK and LH reviewed and verified the data in the study.

Funding

This study was supported by Hebei Natural Science Foundation, No.H2024206504.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of The Fourth Hospital of Hebei Medical University (approval no. 2024KY108). Conducted in line with institutional ethics and the Declaration of Helsinki principles, it involved no human subject interactions or use of identifiable information. Thus, informed consent for using existing pathology materials was waived, and patient privacy was strictly protected.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.