Breast cancer is the most commonly diagnosed cancer among women worldwide, posing a significant global health burden (1). The prognosis of breast cancer patients is closely tied to the stage of disease progression, with survival rates declining sharply upon the development of distant metastases. As a result, early detection is critical to improving therapeutic outcomes and patient survival (2, 3).

The evaluation of metastatic status in breast cancer patients involves both preoperative and postoperative assessments. Postoperative analyses provide detailed information about tumor characteristics, including lymphatic and vascular invasion, the extent of necrosis, and the degree of epithelial hyperplasia (4). While these parameters are valuable indicators of metastatic potential, they are only available after surgical intervention. Preoperative diagnostic tools, such as imaging modalities, offer preliminary insights into tumor behavior. However, definitive assessment of lymph node involvement—a key determinant of metastatic spread—relies on the analysis of the sentinel lymph node (SLN) obtained during surgery (5, 6) ( Figure 1A ). SLN biopsy is a minimally invasive technique that enables early detection of tumor cells disseminated from the primary site, facilitating precise therapeutic planning and reducing the need for more extensive surgical procedures (4, 5).

Following SLN biopsy, the tissue is processed and evaluated by pathologists through microscopic examination. Tissue structures are differentiated using staining techniques, such as hematoxylin and eosin (HE) for general morphology or immunohistochemistry (IHC) for specific biomarkers (7) ( Figures 2A, B , respectively). IHC is particularly valuable for identifying metastatic cells within normal parenchyma, especially in cases where morphological differentiation between normal and abnormal tissue is challenging using HE staining alone (8, 9) ( Figure 1B ). The IHC process involves the use of primary antibodies that bind to specific antigens in the tissue, followed by enzyme-conjugated secondary antibodies that produce a visible signal, typically a brown precipitate, at the target site (10).

Advances in medical technology have enabled the digitization of histopathological slides, resulting in whole-slide images (WSIs). WSIs capture multiple magnifications of the same tissue region within a single file, allowing pathologists to navigate seamlessly from a macroscopic overview to high-resolution details. This digital transformation has facilitated the application of computer vision algorithms, particularly convolutional neural networks (CNNs), to automate the detection of tumor regions in biopsy specimens. Researchers have developed CNN-based models capable of identifying various cancer types, including breast cancer, within WSIs (11–17). These models are trained on annotated datasets containing both tumor and healthy tissue samples, enabling them to learn discriminative patterns for accurate classification (18). To manage the computational demands of processing large WSIs, images are often divided into smaller patches at the highest resolution. The CNN generates patch-wise predictions, which are subsequently aggregated to produce comprehensive diagnostic visualizations (19) ( Figure 1C ).

Despite its potential, computational pathology faces two key challenges. First, the limited availability of patient data often requires multi-institutional collaborations, delaying research progress (20). Second, the computational resources needed for training models are immense, with studies requiring terabytes of storage and hundreds of thousands of GPU hours (21). These demands strain infrastructure and contribute to environmental concerns, as medical imaging accounts for approximately 1% of global greenhouse gas emissions (22).

This study proposes a transfer learning approach to address critical challenges in computational pathology for metastatic tumor detection. Building on the established success of deep learning in medical imaging, we investigate whether pretrained models can be effectively adapted to overcome data scarcity and computational constraints in clinical settings. Our work is guided by three principal hypotheses: First, that transfer learning with strategic fine-tuning can extract diagnostically relevant patterns from limited histopathology datasets while maintaining clinical-grade accuracy. Second, combining institutional data with curated public repositories will yield more robust feature representations than either source alone. Third, integrating intuitive visual interpretability tools can render model decisions transparent to pathologists without requiring specialized technical expertise.

The experimental framework employs a ResNet50 architecture initialized with ImageNet weights, subsequently fine-tuned on annotated whole-slide images of sentinel lymph node biopsies from Hospital Clinic Barcelona ¹ . Through carefully designed comparisons across eight experimental conditions, we systematically evaluate the impact of key factors including: (i) data composition strategies balancing local and public datasets, (ii) augmentation techniques for improved generalization, and (iii) evaluation protocols that assess both computational and clinical viability.

A novel aspect of this research is its dual-focus evaluation methodology, which combines traditional performance metrics with clinician-centric visualization tools. This approach seeks to establish whether computational predictions can simultaneously satisfy quantitative benchmarks and operational requirements for pathological diagnosis. The study design intentionally avoids presupposing optimal configurations, instead examining how varying degrees of data diversity and explainability influence model behavior in diagnostically relevant scenarios.

The detection of metastasis from primary tumors is a critical step in tumor staging, with SLN biopsies serving as the gold standard for assessing lymph node involvement (5, 6). Historically, pathologists relied on optical microscopes to examine processed and stained tissue samples (23). However, the introduction of digital slide scanners has transformed this process, enabling the creation and storage of WSIs that can be analyzed computationally (24).

A central challenge in automated WSI analysis is the accurate classification of tumor regions. This requires the generation of masks—annotations that distinguish cancerous cells from healthy tissue. Masks can be binary, indicating the presence or absence of cancer, or multi-label, providing additional information such as malignancy grades or the presence of immune cells within the tumor microenvironment (25, 26). These masks are typically created by pathologists or derived from ICH techniques. In the latter approach, IHC is used to identify cancer regions in a tissue slice adjacent to the HE-stained slide. The spatial information from the IHC slice is then transferred to the HE slide to generate the corresponding mask (26, 27).

In machine learning-based analysis of WSIs, the images are typically divided into smaller patches, where patch size serves as a critical parameter affecting both computational efficiency and the model’s capacity to capture relevant contextual information (19, 28, 29). To mitigate challenges associated with limited dataset size and variability, data augmentation techniques—including rotation, contrast adjustment, noise injection, and grayscale conversion—are widely adopted (15, 16, 30, 31). While such augmentations can enhance model robustness and generalization, excessive application may introduce artificial patterns or distort underlying data distributions (32). Notably, some studies mention some benefits of augmentation without providing full comparative analyses against non-augmented baselines, raising questions about their necessity in certain contexts (33).

Two primary methodologies are commonly employed for training models in WSI patch classification. The first is transfer learning, wherein a model pretrained on a large-scale dataset (e.g., ImageNet) is fine-tuned for the target histopathological task. A key variant of this approach is end-to-end learning, which unifies feature extraction and classification within a single framework, enabling direct learning from image patches. This method is particularly beneficial when working with limited annotated data, as it leverages pre-existing feature representations. Widely used architectures for transfer learning include ResNet, AlexNet, DenseNet, EfficientNet, Inception V3, and VGG (34–37). For example, ResNet-50 has demonstrated strong performance in breast cancer histopathology detection (38). However, a notable limitation of these deep learning models is their inherent “black-box” nature, often lacking interpretability in decision-making processes.

The second methodology separates feature extraction from classification. Handcrafted feature extraction involves using domain specific knowledge and algorithmic methods to identify and quantify relevant features within an image. These features are then fed into a separate classifier, such as a Support Vector Machine (SVM) or Random Forest (34, 39, 40). Examples of handcrafted features include those derived from texture analysis (e.g., Local Binary Patterns, Gabor filters), edge detection (e.g., Canny edge detector), or feature descriptors (e.g., SIFT, HOG) (41–43). While handcrafted features are interpretable, they require extensive domain expertise and may not capture high-level abstractions as effectively as deep learning models.

The performance of deep learning models in histopathology is typically evaluated using a combination of quantitative metrics and qualitative visualizations. Standard quantitative measures include accuracy, sensitivity, specificity, precision, recall, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) (19, 28, 40). On the qualitative side, visual assessments often involve the use of heatmaps, which highlight regions of interest within WSIs. These heatmaps can be displayed independently or overlaid on the original tissue images, occasionally requiring grayscale conversion to improve visual clarity (28, 44). Such visualizations play a critical role in aiding medical professionals by enhancing the interpretability of model predictions and by facilitating the identification of cancerous regions, which in turn can support the development of new training datasets (28).

Among the various visualization techniques, Class Activation Map (CAM)-based methods are widely used for explaining CNN outputs. These techniques work by projecting back the weights from the final convolutional layers to generate saliency maps that highlight spatial regions most influential to a model’s prediction (45). While effective for small-scale images, applying CAM-based methods to WSIs—which can exceed 10 billion pixels at high resolution—poses significant challenges. First, generating maps for each of the thousands of patches in a WSI can be computationally expensive. Second, the resulting pixel-level heatmaps across such a massive canvas can become visually overwhelming, introducing excessive detail that obscures broader diagnostic patterns. Third, in contrast to natural images where objects of interest are often visually prominent, regions of interest in pathology images tend to be subtle and lack strong saliency or clear boundaries. This makes it difficult for CAM-based methods to consistently localize relevant features, often resulting in under- or over-activation of regions and reducing the interpretability of the visualization at the clinical scale (45–47).

In summary, despite notable advancements in automated WSIs analysis, key challenges persist in addressing data variability, optimizing data augmentation, and improving model interpretability. Our work advances current methodologies by (1) incorporating external datasets to better capture domain variability, (2) systematically evaluating the influence of augmentations on model performance, and (3) developing visualization tools to increase the transparency of deep learning models in metastasis detection.

In this section, we present the results of our experiments, as summarized in Table 2 . Initially, we computed histograms for the 4 patch datasets generated. Supplementary Figure 1 illustrates that the histograms for positive samples in all 4 patch datasets exhibited a similar distribution to those of the negative samples, confirming an expected increase in the relative frequency of pixels with value 255 due to white patches in negative samples.

In the subsequent fine-tuning and patch-based testing phase, both processes utilized patches derived from the same WSIs. As detailed in Table 4 , optimal performance across all models was consistently achieved when fine-tuning was performed without the application of data augmentations. Specifically, the ‘Interhospital’ model (Experiment 7) demonstrated the highest accuracy (0.881), while the ‘Tumor + FIB’ model (Experiment 3) exhibited the best sensitivity (0.756) and weighted F1-score. Notably, the ‘Tumor’ model (Experiment 1) achieved the highest specificity (0.987). Regarding computational efficiency, model training times, conducted using NVIDIA Tesla K80 hardware for a fixed 18 epochs per dataset, varied considerably, and this variation appears correlated with dataset size. The ‘Tumor’ and ‘Tumor + FIB’ datasets, comprising 23,900 samples each, showed the most rapid training, both completing in approximately 5.87 minutes. In contrast, the ‘Interhospital’ dataset, with 104,107 samples, required a significantly longer duration of 25.55 minutes, and the ‘Tumor extended’ dataset, the largest with 227,923 samples, presented the longest training time at 55.93 minutes.

Following the fine-tuning phase, our research extended to the whole slide testing phase ( Table 2 , Whole Slide Testing). The metrics obtained during this phase differed significantly from the preliminary patch-based testing metrics ( Table 4 ). To capture these differences, two tables were generated. Table 5 outlined the results for each fine-tuned model with and without augmentations, while Table 6 detailed the results for each image (individual patient) across all models, both with and without augmentations. These analyses aimed to identify robust models across all images and assess the difficulty levels in detecting metastasis for each image using all models.

Upon analysis of this testing phase, models fine-tuned with augmentations consistently outperformed their non-augmented counterparts, as evidenced by the results in Table 5 . Particularly, the ‘Interhospital’ model (Experiment 8) achieved the highest accuracy (0.982), sensitivity (0.256), and F1-score (0.978). Additionally, ‘Tumor Extended’ (Experiment 6) demonstrated the highest specificity (0.998). The Area Under the ROC Curve (AUC-ROC) further supported ‘Interhospital’ as the leading model ( Supplementary Figure 2 ).

Examining patients individually, superior metrics were also evident in cases where augmentations were utilized ( Table 6 ), aligning with the trends previously observed in the whole slide testing phase ( Table 5 ). This observation emphasized the potential impact of augmentations on model performance, especially in scenarios involving domain changes and unseen images. Notably, patients 16 and 18 achieved the best accuracy and F1-score metrics, despite having low tumor patch content in their WSIs (0.03% and 0.35%, Table 3 ). This outcome indicated the models’ proficiency in identifying healthy tissue, offering potential benefits in the screening of patients with normal SLN biopsies.

Quantitative metrics analysis revealed that superior metrics were achieved without augmentations during model fine-tuning and patch-based testing ( Table 4 ). This result was likely due to the low variability between the training and testing sets, sourced from the same WSIs. However, during whole slide testing, where domain changes occurred, models trained with augmentations demonstrated better performance ( Tables 5 , 6 ). This underscored the importance of using models trained on diverse data when transitioning domains, enhancing their generalization capabilities.

The qualitative metrics section served as a vital complement to the quantitative findings, offering valuable clinical insights. Cases with minimal tumor pixels, such as cases 16 and 18 ( Figure 5 ), displayed distinct patterns from those with higher tumor counts. These cases exhibited tumor pixel rates of 0.03% and 0.35% across their WSIs, respectively ( Table 3 ). The top-performing model, Interhospital, showed intersections with tumor regions in its predictions, but false positives were scattered throughout the images. Notably, false positives were also observed above areas of heterogeneous tissue with some degree of fibrosis in case 18 ( Figure 5 ).

Furthermore, we evaluated the models’ performance in cases featuring a substantial proportion of tumor pixels per image. Specifically, we analyzed WSI case 25 (depicted in the fourth and fifth rows) and case 27 (shown in the final row of Figure 5 ).

These cases displayed tumor pixel ratios of 18.64% and 5.11%, respectively (refer to Table 3 ). A significant increase of at least 1600% in pixel percentage was observed when comparing cases with low and high tumor pixel percentages. This considerable rise in tumor pixels correlated with the tumor cell count, indicating an extended period of tumor growth from the initial metastatic cell invasion of the SLN to the biopsy. Consequently, this prolonged duration likely contributed to the distinct differences between tumor cells and normal cells visual phenotype, improving the model’s performance.

In the context of WSI 25, both tissue slices exhibited striking similarities between the ground truth area for IHC and the model’s predictions, as demonstrated in Figure 5 cases 25.1 and 25.2. Conversely, in WSI 27, while the IHC and binary predictions showed a significant overlap (highlighted in red), the color-coded prediction map provided valuable insights ( Figure 5 case 27). Specifically, regions with the highest predicted probabilities (0.9 and above, indicated in red) almost entirely coincided with tumor areas. The remaining predictions, characterized by lower probabilities (below 0.9, depicted in green and yellow), played a vital role in distinguishing false positives from true positives (highlighted in red). This detailed analysis underscored the model’s accurate identification of tumor regions in cases with varying tumor pixel percentages, offering crucial information for clinical evaluation and diagnosis.

Moreover, upon a meticulous inspection of the prediction maps generated by all models overlaid on the same image ( Figure 6 ), the distinctive outcomes align with the quantitative metrics, as illustrated in the F1 score results in Table 5 and the AUC ROC values presented in Supplementary Figure 2 . Notably, the ‘Interhospital’ model consistently outperforms other models across all three scenarios, followed by ‘Tumor + FIB’. The preceding analysis provides two critical insights. First, it highlights the importance of the visual interface for comprehensively evaluating model performance. The interface enables direct visualization of model outputs in the context of the original histopathology images, conferring clinical meaning to quantitative metrics. Second, it underscores the significance of heterogeneity in the whole slide testing. As observed, the top performing models were fine-tuned on the most heterogeneous datasets. Specifically, the ‘Interhospital’ model was trained on samples from two distinct sources - locally generated data and the public Camelyon dataset. Meanwhile, the ‘Tumor + FIB’ model was trained on patches containing not only tumor cells, but also associated fibrous stromal tissue.

This study demonstrates the effectiveness of transfer learning in fine-tuning ResNet50 models for accurately detecting metastatic regions in SLN biopsy images. By leveraging Cytokeratin markers in IHC images to generate masks, we created four distinct patch datasets for fine-tuning, incorporating both augmented and non-augmented samples. Patch-level testing revealed that the highest accuracy was achieved without augmentations, likely due to overlaps between test and training samples from the same WSIs. However, whole-slide testing showed improved performance with augmentations, underscoring the role of data variability in enhancing model generalization. Notably, incorporating additional variability—such as introducing samples from Camelyon into the Interhospital dataset—provided the optimal diversity level, leading to superior performance on unseen images. These findings suggest that supplementing data augmentation with diverse datasets that maintain domain specificity can further enhance model robustness.

Our results provide support for our three initial hypotheses. First, transfer learning with appropriate fine-tuning indeed achieved clinically relevant performance (accuracy up to 0.982) despite using only 25 WSIs for training, supporting that pretrained models can overcome data scarcity limitations. Second, the hybrid Interhospital dataset consistently outperformed models trained on single-institution data (accuracy improvements of 3-5%), demonstrating that combining institutional with public datasets enhances generalization. Third, our visual explainability tools were able to bridge the technical-clinical gap - pathologists could interpret the heatmaps and binary maps with minimal training, validating their utility in real diagnostic workflows. These findings collectively help demonstrate that our proposed approach addresses the key challenges of data efficiency, generalizability, and clinical adoption in computational pathology.

The observed variations in model training times reveal important insights about computational efficiency. Models trained on smaller datasets (‘Tumor’ and ‘Tumor + FIB’ with 23,900 samples) completed training significantly faster than those using larger collections. The ‘Interhospital’ dataset (104,107 samples) showed moderate training duration, while the extensive ‘Tumor Extended’ dataset (227,923 samples) required the longest processing time, confirming the expected scaling of computational most accurate model (patch-based testing: 0.881; whole-slide: 0.982 with augmentation), emphasizing that data diversity and quality outweigh sheer volume for metastatic detection tasks. This finding has important practical implications for clinical implementation, where both accuracy and computational efficiency are paramount.

We also noted a difference in performance between evaluating individual patches and entire slides. While the model consistently showed high specificity, meaning it was good at correctly identifying non-tumor samples, its sensitivity decreased when analyzing whole slides. This indicates a potential difficulty in detecting new tumor regions not seen during training. Data augmentation proved useful, as models trained with augmented data showed better tumor detection without sacrificing specificity. The best accuracy we achieved (0.98) was with an augmented model, which is comparable to results from prior studies classifying similar lymph node samples using GoogleNet (0.98), AlexNet (0.92), VGG16 (0.97), and FaceNet (0.96) (54–57). Other research on different types of oncology images, also using transfer learning and augmentations (including techniques like edge detection we didn’t explore), reported similar (VGG 19, ResNet 50, ResNet 152, MobileNet) and even higher accuracies (VGG 16, ResNet 101V2, DenseNet 169) (58).

Our evaluation metrics systematically assessed model performance across different testing conditions. When tested on previously unseen WSIs, models exhibited overall accuracy improvements, likely due to the higher proportion of non-tumor regions in WSIs, which enabled models to leverage their high specificity. However, the observed decrease in sensitivity highlights ongoing challenges in generalizing to new tumor data. The increase in weighted F1-score with additional training data suggests that models benefit from the larger proportion of healthy tissue in whole-slide images compared to tumor-enriched patch datasets. The ability to reliably identify non-tumor regions could facilitate computational pathology workflows by enabling automatic masking of healthy areas, allowing pathologists to focus on uncertain regions.

A key aspect of this work was creating a visual interface to improve clinical understanding. Overlaying identified metastatic areas onto the complete patient slides allowed pathologists to visually evaluate the model’s predictions within the original context. Our results demonstrated accurate tumor patch detection when tumor content was above 16%, consistent with our quantitative findings. Indeed, our patch-based heatmap overlay offered the ideal level of detail for pathologists to differentiate zones of tumor development in the predictions. As noted, larger metastatic deposits tend to exhibit greater cellular diversity due to longer growth periods and potential variations in oxygen and nutrient access, leading to distinct morphologies within the same cluster (59).

Additionally, our approach offers a time-efficient alternative to traditional IHC procedures. Generating IHC-stained slides from biopsy tissue can take up to 16 hours (60, 61), whereas our models process an HE-stained WSI in approximately 18 minutes—a 98% reduction in processing time. By enabling rapid, automated identification of tumor and healthy regions, this approach has the potential to streamline computational pathology workflows and reduce clinician workload. However, challenges persist in cases with low tumor cell content or poorly differentiated morphology—such as Cases 16 and 18—where tumor cells closely resemble normal tissue. In addition to the lack of clear differentiation, these tumor regions covered a relatively small surface area, resulting in fewer representative patches with this morphology in the training data. This raises an important question: is the issue primarily due to poor differentiation, or is it the limited representation of such cases in the training dataset? To explore this, future studies should focus on datasets containing a sufficient number of poorly differentiated tumor samples alongside normal tissue. Despite these limitations, the model’s ability to accurately exclude healthy tissue and flag uncertain regions for expert review highlights its time-saving potential compared to manual IHC slide analysis.

The clinical implications of this work extend beyond technical performance metrics. Our framework demonstrates how computational pathology can be strategically adapted to institutional settings through three key principles: leveraging transfer learning to overcome data limitations, curating hybrid datasets that balance domain specificity with diversity, and developing interpretation tools that align with pathologists’ diagnostic workflows. This approach not only achieves diagnostic accuracy comparable to manual assessment but does so while respecting the practical constraints of clinical environments, where computational resources and annotation bandwidth are often limited.

Looking forward, two parallel pathways emerge for advancing this research. First, technical refinements should focus on improving detection of diagnostically challenging cases, particularly micrometastases and tumors with ambiguous morphology. Second, clinical validation studies must establish how best to integrate such systems into real-world workflows, including optimal division of labor between AI and pathologists. The interpretability tools developed here provide a foundation for this transition, enabling collaborative decision-making where computational efficiency complements human expertise. Together, these directions promise to accelerate the translation of AI-assisted pathology from research laboratories to routine clinical practice.

This study validates three fundamental advancements in computational pathology for metastatic detection. First, we demonstrate that transfer learning with targeted fine-tuning can achieve clinical-grade accuracy using limited training data, overcoming establish that intuitive visualization tools effectively bridge the gap between computational outputs and clinical interpretation, facilitating pathologist adoption.

The success of our framework stems from its synergistic approach: (1) optimizing data efficiency through strategic transfer learning, (2) improving robustness via multi-source dataset integration, and (3) ensuring clinical relevance through interpretable visual analytics. These innovations collectively address the tripartite challenge of data scarcity, domain generalization, and workflow integration that has hindered widespread adoption of AI in pathology.

While demonstrating significant improvements over conventional methods, two key opportunities emerge for future research: enhancing detection of minimal residual disease and improving performance on morphologically ambiguous cases. Clinical translation of this validated pipeline could transform metastatic screening protocols by combining the efficiency of AI with pathological expertise, potentially establishing new standards for rapid, accurate lymph node assessment. The principles developed here may extend to other histopathological applications where data limitations and clinical interpretability remain persistent challenges.