This section provides a brief literature review of what technology has been used for the detection of breast cancer over the past few years for breast cancer detection. And this is going to justify our innovative model and method in the further sections.

AI has revolutionized the healthcare landscape, offering transformative capabilities in automating and enhancing diagnostic processes. In the context of breast cancer detection, AI models have demonstrated superior performance in analyzing complex MRI datasets, identifying patterns that may elude human experts (Lehman et al., 2019). For instance, deep learning convolutional neural networks (CNNs) and attention mechanisms have been applied to segment breast tissues, classify lesions, and predict malignancy with high accuracy (Litjens et al., 2017). These advancements not only reduce errors but also minimize the workload of radiologists, making AI a valuable tool in clinical settings. The integration of AI into breast cancer detection workflows has also been shown to address the challenges of class imbalance and variability in imaging quality. Many studies have employed transfer learning and ensemble techniques to overcome these challenges, achieving higher sensitivity and specificity than traditional methods (McKinney et al., 2020). Moreover, AI-driven algorithms can process datasets quickly and consistently, offering significant advantages in population-based breast cancer screening programs (Rodrigues et al., 2021).

Kaymak et al. (2017) applied neural networks to classify 176 histopathology images, using discrete Haar wavelets for preprocessing. Radial Basis Function Networks outperformed Back Propagation Networks, achieving 70.49% accuracy. Fu et al. (2022) predicted Peripherally Inserted Central Catheter (PICC)-related thrombosis in breast cancer patients using Artificial Neural Network (ANN) and Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance, outperforming logistic regression (AUC: 0.742 vs. 0.675). Punitha et al. (2021) combined Artificial Immune Systems and Bee Colony optimization for feature selection, improving multilayer perceptron (MLP) performance on the Wisconsin Breast Cancer Dataset (WBCD) dataset for automated diagnosis.

Sharma et al. (2024) used the WBCD to build a stacked ensemble framework integrating Decision Tree, AdaBoost, Gaussian Naive Bayes, and MLP classifiers, achieving 97.66% accuracy. Ensemble learning and feature engineering contributed to its robust performance. Liu et al. (2024a) enhanced a Visual Geometry Group 16-layer network (VGG16)-based model with transfer learning and focal loss to handle class imbalance in 7,841 mammograms, yielding a 96.95% accuracy. Li et al. (2024b) proposed a multimodal fusion model that combined ResNet34-extracted MRI features with RNA-seq gene expression data, using transformers and attention mechanisms to predict treatment response in breast cancer patients more accurately than single-modality approaches.

Munshi et al. (2024) introduced a novel breast cancer detection method by combining an optimized ensemble learning framework with explainable AI (XAI). Using the Wisconsin Breast Cancer Dataset, which contains 32 numerical features, they employed an ensemble model integrating CNNs with traditional machine learning algorithms like random forest (RF) and support vector machine (SVM). U-NET was used for image-based tasks, and a voting mechanism combined RF and SVM predictions. The model achieved an impressive accuracy of 99.99%, surpassing state-of-the-art methods in precision, recall, and F1-score. The integration of XAI enhanced the interpretability and transparency of model decisions, further improving breast cancer diagnostics.

Atrey et al. (2024) developed a multimodal classification system for breast cancer by integrating mammogram (MG) and ultrasound (US) images using deep learning and traditional machine learning techniques. The dataset, sourced from AIIMS, Raipur, India, consisted of 31 patients and 43 MG and 43 US images, which were augmented to 1,032 images. Data preprocessing involved manual region-of-interest (ROI) annotation by radiologists and noise filtering techniques for both image types. The deep learning model, ResNet-18, was used for feature extraction, while SVM classified the fused features. The hybrid ResNet-SVM approach achieved a classification accuracy of 99.22%, significantly outperforming unimodal methods. This study demonstrates the effectiveness of combining deep learning with traditional machine learning for improved breast cancer diagnosis.

Robinson and Preethi (2024) developed the SDM-WHO-RNN model, combining fully convolutional networks with RNNs for classifying 9,109 histopathological images from the Breast Cancer Histopathological Image (BreaKHis) and BH datasets. Preprocessing steps included resizing, noise reduction, and stain normalization, while Least Squares Convolutional Encoder-Decoder (LS-CED) was used for tumor segmentation. The system achieved 97.9% accuracy in detecting complex tumor cells. Zhang et al. (2024) proposed the Convergent Difference Neural Network (V-CDNN), an ensemble of Convergent Difference Neural Networks optimized with a Neural Dynamic Algorithm and softsign activations. Trained on public datasets, the model reached a perfect 100% classification accuracy, demonstrating both speed and robustness.

In their 2021 study, Meha Desai and Manan Shah compared the effectiveness of MLP and CNN for breast cancer detection. The study used datasets such as BreakHis, WBCD, and WDBC, with images and biomarkers preprocessed through normalization and feature extraction (e.g., Discrete Cosine Transform). While both MLP and CNN were trained to classify breast abnormalities, CNN consistently outperformed MLP in terms of accuracy. CNN achieved ~99.86% accuracy on the BreakHis dataset, while MLP’s performance was generally lower. CNN was found to be more reliable for image-based diagnosis, whereas MLP showed limitations for larger datasets.

Further to this, Ezzat et al. (2023) proposed an optimized Bayesian convolutional neural network (OBCNN) for detecting invasive ductal carcinoma (IDC) from histopathology images. This approach integrated ResNet101V2 with Mucinous Carcinoma (MC) for uncertainty estimation. Using the slime mould algorithm to optimize dropout rates, the study fine-tuned the architecture, outperforming other pre-trained networks such as VGG16 and DenseNet121. The model demonstrated significant robustness in diagnosis and generalization. Joseph et al. (2022) proposed a multi-classification approach for breast cancer using handcrafted features, including Hu moments, Haralick textures, and color histograms. These features were combined with deep neural networks trained on the BreakHis dataset. Data augmentation was applied to address overfitting, and a four-layer dense network with softmax activation was employed to enhance classification accuracy.

Paul et al. (2023) proposed a novel breast cancer detection system using an SDM-WHO-RNN classifier combined with LS-CED segmentation. Their approach incorporated preprocessing steps, such as noise elimination and image normalization. The LS-CED segmentation localized nuclei, which were then classified based on size and shape. The system demonstrated high accuracy in detecting cancerous cells. Taheri and Omranpour (2023) introduced an ensemble meta-feature generator (EMFSG-Net) for classifying ultrasound images. Leveraging transfer learning with the VGG-16 architecture, the model employed support vector regression (SVR) to create efficient feature spaces. To address issues like overfitting and dead neurons, the leaky-ReLU activation function was incorporated, leading to improved feature representation and classification performance. Muduli et al. (2021) proposed a deep CNN for automated breast cancer classification using mammograms and ultrasound images. The model, consisting of five learnable layers, employed manual cropping for feature extraction and data augmentation to improve generalization. The approach achieved superior results compared to state-of-the-art methods.

Dr. Suvidha Tripathi’s research contributes significantly to the field of breast cancer histopathology image classification, particularly through innovative approaches that integrate spatial context and hybrid feature representations. In one study, Tripathi et al. proposed a BiLSTM-based patch modeling framework, which treats histopathology image patches as sequential data to capture spatial continuity—an aspect often overlooked in conventional CNNs. This model demonstrated strong performance on the Breast Cancer Histology (Grand Challenge 2018 dataset), highlighting the value of contextual learning in improving classification accuracy. In subsequent work, she introduced a hybrid architecture combining CNNs with Bag-of-Visual-Words (BoVW), effectively integrating handcrafted and deep features to enhance discrimination in limited data settings. This approach outperformed standard deep networks like ResNet and DenseNet on the same dataset, underscoring the benefits of feature selection and hybrid modeling in medical imaging tasks. Together, these studies emphasize alternative paths for improving classification performance—either by modeling inter-patch relationships or by enriching feature spaces—both of which are highly relevant for advancing automated diagnosis in breast cancer histopathology.

This research is toward an innovative approach for breast cancer detection, including the multiple subclasses of breast cancer, which still confuses medical doctors, and it hence makes patients undergo more advanced invasive tests for a better understanding of the disease. The technique is based on an AI model trained on a sizable batch of data. Section 2 discusses the model in detail, Section 3 shows the Results and Discussion, and the Conclusion is given in Section 4.

In this study, we utilize the BreaKHis dataset compiled by Spanhol et al. (2016), which contains 9,109 microscopic images of breast tissue captured at four magnification levels: 40×, 100×, 200×, and 400×. These images are categorized into two overarching classes—benign and malignant tumors. Within each class, the dataset further identifies four subtypes: for benign tumors, these include adenosis (A), fibroadenoma (F), phyllodes tumor (PT), and tubular adenoma (TA); and for malignant tumors, ductal carcinoma (DC), LC, mucinous carcinoma (MC), and papillary carcinoma (PC). The image filenames in the BreaKHis dataset encode multiple details about each sample, including the biopsy method used, the tumor’s classification (benign or malignant), its histological subtype, the patient ID, and the magnification level. For instance, the file named SOB_B_TA-14-4659-40-001.png refers to the first image of a benign tumor classified as TA, obtained from patient 14-4659 at a magnification of 40× using the Segmental Orthogonal Biopsy (SOB) technique (Spanhol et al., 2016).

It comprises samples from 82 patients collected at a single institution in Brazil, with no publicly available metadata on age, ethnicity, or receptor status—24 with benign tumors and 58 with malignant tumors. Of the 7,909 images used in the final analysis, 5,429 are malignant and 2,480 are benign, distributed across the aforementioned magnification levels. Following preprocessing and class balancing, the revised class distribution is outlined in Table 1.

Benign tumors, by histological standards, lack features associated with malignancy, such as cellular atypia, mitotic activity, invasive growth, or metastatic potential. They typically exhibit localized, slow growth and are considered non-lethal (Robbins and Cotran, 2010). In contrast, malignant tumors are characterized by their capacity to invade neighboring tissues and metastasize to distant sites, often posing life-threatening risks (Robbins and Cotran, 2010). All tissue samples in this dataset were acquired through SOB—a surgical procedure also referred to as partial mastectomy or excisional biopsy. Unlike needle-based techniques, SOB yields larger tissue specimens and is generally performed under general anesthesia in a clinical setting (American Cancer Society, 2019). The classification of tumors into distinct subtypes is based on their cellular morphology under microscopic examination, which holds significance for both prognosis and treatment planning (National Cancer Institute, 2021).

The BreakHis dataset was used for both binary (benign vs. malignant) and multiclass subclass classification. All histopathological images were resized to 128 × 128 pixels and normalized to a standard RGB range. A stratified 80:20 train-test split was employed to preserve class balance across sets. For cross-validation, a 5-fold split was used with data shuffling to ensure unbiased performance evaluation, and no image was shared across folds.

Prior to splitting, real-time augmentation was applied separately to the benign and malignant image sets, increasing the number of benign samples to 7,440 and malignant samples to 9,385. The augmented dataset, totaling 16,825 images, was then split into training, validation, and test sets using a stratified class-wise approach. For both benign and malignant classes, 20% of the data was reserved for testing, and the remaining 80% was further divided in a 75:25 ratio to form the training and validation sets. This resulted in 10,095 images in the training set and 3,365 images each in the validation and test sets.

To improve generalization and address the risk of overfitting due to limited data, we implemented an extensive on-the-fly augmentation pipeline using TensorFlow’s data generators. This included spatial and color transformations to simulate natural histological variation while preserving key tissue features. Table 2 summarizes the augmentation parameters used during training.

These augmentations were applied to each batch during training to expose the model to diverse variations without expanding the dataset size. In addition, we used early stopping (patience = 5) and ReduceLROnPlateau scheduling to prevent overfitting, along with dropout (0.45), L2 kernel regularization, and batch normalization in each dense layer. Learning curves were generated during each fold to monitor training behavior. The curves showed a minimal gap between training and validation accuracy, indicating stable convergence and effective generalization across classes.

To address class imbalance—particularly for underrepresented subtypes, such as PT and LC—we applied targeted data augmentation during preprocessing. For most classes, each image was augmented using 90° and 180° rotations along with horizontal flipping, effectively tripling their sample size. However, for class label 4 (DC), which was already sufficiently represented, only minimal augmentation was applied to prevent overrepresentation. This strategy helped balance the dataset and reduce bias during training, without altering the loss function or applying weighted sampling.

A CNN is a deep learning architecture tailored for grid-like data such as images, leveraging spatial hierarchies to extract local and global features (LeCun et al., 1998). Input images are processed through convolutional layers that apply learnable filters to detect features like edges and textures, generating spatially preserved feature maps (Krizhevsky et al., 2012). Non-linear activation functions, typically ReLU (Glorot et al., 2011), introduce complexity, while pooling layers reduce spatial dimensions and enhance translational invariance (Scherer et al., 2010). Deeper layers capture higher-level abstractions, which are then passed to fully connected layers for classification (Simonyan and Zisserman, 2014). Enhancements such as batch normalization, residual connections (He et al., 2016), and attention mechanisms have further advanced CNN performance, particularly in medical image analysis (see Figure 2).

The proposed classification algorithm is distinguished by its multi-scale feature fusion architecture built on DenseNet121. Instead of relying only on the final layer features (as done in most standard CNN classifiers), it extracts and fuses features from three different depths of the DenseNet (the end of conv3, conv4, and conv5 blocks). Each feature map is subjected to GAP and L2 normalization, then passed through a small fully connected layer with batch normalization (and dropout) before concatenation. This design enables the model to capture discriminative patterns at multiple spatial scales—from fine-grained cellular details to higher-level tissue organization—within a single network. In essence, the model behaves like an integrated multi-scale ensemble, where lower-level texture features and higher-level semantic features jointly inform the final prediction. This multi-scale fusion is particularly well-suited to histopathology images, where diagnostically relevant cues appear at different magnifications. Prior work has recognized the value of multi-scale information: Araújo et al. (2017) designed a custom CNN to explicitly capture both nuclei-level details and global tissue architecture (Araújo et al., 2017), and Gupta and Bhavsar (2018) similarly argued that “different layers… contain useful discriminative information” for classifying breast histology images (Gupta and Bhavsar, 2018). The proposed model builds on this principle by simultaneously leveraging DenseNet’s intermediate feature maps, rather than using only the deepest features or requiring multiple separate networks.

By fusing intermediate feature vectors, our approach yields a richer representation that combines context from multiple receptive field sizes. This strategy is analogous to the “hypercolumn” or feature pyramid concept—effectively consolidating features at various levels of abstraction. In histopathology, such multi-level fusion has clear advantages: for example, benign vs. malignant differentiation may depend on both cellular morphology and the broader tissue architecture. Traditional transfer-learning classifiers often discard these multi-level cues by only using the final global feature map of a pre-trained CNN. In contrast, our DenseNet121-based model preserves multi-scale feature information by tapping conv3, conv4, and conv5 outputs. Zhu et al. (2019) demonstrated a similar idea with a two-branch “global vs. local” CNN ensemble, where merging a whole-image branch with a high-zoom patch branch improved representation power (Zhu et al., 2019). Our method achieves a comparable multi-scale effect within a single backbone: the conv3_block output (earlier layer) can focus on local textural patterns (analogous to a high-magnification view), while the conv5_block output provides global context (analogous to a low-magnification overview). The DenseNet architecture’s dense connectivity may further facilitate this feature reuse across scales (Huang et al., 2017). Notably, DenseNet121 has been a popular choice in medical image analysis due to its efficient feature propagation; several breast histopathology studies report strong results with DenseNet-based transfer learning—for example, Liew et al. (2021) achieved ~97% accuracy by using DenseNet201 features with an XGBoost classifier (Liew et al., 2021). Our model differentiates itself by not only fine-tuning DenseNet121 on the task, but by modifying its topology to combine multi-scale features—a design rarely explored in prior breast cancer studies.

Another novel aspect of our approach is the use of L2 normalization and balanced regularization in the feature fusion process. Each branch’s pooled feature vector is L2-normalized before fusion, ensuring that no single scale’s features dominate the others due to scale differences. This is an uncommon yet effective practice in classification networks—more frequently seen in metric learning or multi-instance aggregation contexts—and it promotes stability when training the fused classifier. For instance, Xie et al. (2020) applied L2 normalization to patch feature vectors in a histopathology model to stabilize feature magnitudes (Xie et al., 2020). In our design, the L2 norm acts similarly to a scale calibration across the three feature streams, so that the subsequent dense layers operate on features of comparable norm. We then apply dropout and batch normalization in each branch’s dense layer (as well as the final classification layer) to combat overfitting. High-resolution histology images are prone to overfitting given their limited datasets, and regularization is critical. Previous works have indeed found that naive CNN models can overfit histopathology data even with dropout—for example, a shallow CNN still overfit BreakHis patches even with 20% dropout, as reported by Kode and Barkana (2023) (Kode and Barkana, 2023). By combining dropout with batch normalization and the inherent regularization of feature fusion, our model aims to generalize better. The inclusion of batch normalization after the small dense layers also helps each scale-specific feature vector to be well-conditioned before merging, which eases the joint learning. These design choices—L2 normalization of features and dropout-regularized dense layers for each scale—are unique refinements that distinguish our architecture from standard fine-tuned CNN classifiers that often use only a global pooling and a linear classifier on top.

Compared to existing deep learning methods in breast cancer histopathology, the proposed model offers a new balance of simplicity and multi-scale sophistication. Many prior studies have focused either on transfer learning with single CNNs or on ensembles of multiple models. On the one hand, simple transfer-learning approaches, fine-tuning a single network (e.g., ResNet or DenseNet), have shown strong baseline performance (Spanhol et al., 2016; Wahab et al., 2017), but they might miss multi-scale cues. Some recent works augment such models with attention mechanisms—for example, an EfficientNetV2 with a channel-spatial attention module (CBAM) by Aldakhil et al. (2024) outperformed standard CNNs, reaching ~92–94% accuracy on BreakHis magnifications (Aldakhil et al., 2024). Others integrate vision transformers to capture global context—for example, Mehta et al. (2022) introduced a ViT-based HATNet that achieved state-of-the-art accuracy (Mehta et al., 2022). However, these sophisticated models often remain single-scale in the sense that they ultimately rely on one level of feature map (global features) enhanced via attention. On the other hand, ensemble and multi-branch strategies explicitly leverage multi-scale inputs. For example, Umer et al. (2022) proposed 6B-Net, a six-branch CNN with different receptive field sizes in each branch, and fused their outputs for classifying the eight BreakHis classes (Umer et al., 2022). Similarly, an earlier study by Zhu et al. (2019) assembled multiple compact CNNs (and even pruned channels) to combine local and global predictions for histology images (Zhu et al., 2019). These methods confirmed that combining features from multiple scales or multiple models can boost accuracy. Our approach achieves this multi-scale fusion within a single pre-trained DenseNet, rather than requiring separate networks for different magnifications or an external ensemble of classifiers. This provides a more unified and computationally efficient framework: for instance, Wakili et al. (2022) reported a DenseNet-based transfer model (“DenTnet”) that attained ~99% accuracy on BreakHis (Wakili et al., 2022), but that model treats DenseNet as a black-box feature extractor for an SVM or softmax classifier. In contrast, our model integrates the multi-level feature extraction into the training process, which could offer better synergy between feature learning and classification. It also simplifies the pipeline—there is no need for post-CNN classifiers like SVM (as used by Araújo et al., 2017 and others) or gradient boosting ensembles (as in Liew et al., 2021) because the fused deep features are learned end-to-end to directly optimize classification.

In summary, the novelty of our algorithm lies in combining multi-scale feature fusion, transfer learning with DenseNet121, and strategic normalization/regularization into one coherent model for breast histopathology classification. This design captures the multi-scale nature of tissue patterns more explicitly than standard CNN classifiers. It leverages DenseNet’s strength in feature reuse while addressing multiple scales akin to multi-branch networks, but with fewer parameters and a single-pass inference. By comparing with the literature, we see that our approach is unique in how it fuses intermediate CNN features: Gupta and Bhavsar (2018) sequentially processed DenseNet layers’ outputs rather than fusing them, and most other DenseNet-based methods (Wakili et al., 2022; Liew et al., 2021) used only the final layer or combined whole-network outputs in ensembles. Moreover, the inclusion of L2-normalized feature vectors and dropout-regularized dense layers for each scale is a novel architectural choice that, to our knowledge, has not been explicitly reported in prior breast cancer histopathology studies. These innovations position our model as a multi-scale, multi-level classifier that is well-aligned with the visual hierarchy a pathologist employs (from cells to tissue architecture), setting it apart from conventional single-scale CNN models and even from recent attention-based or ensemble-based state-of-the-art methods in this domain.

In AI, especially in domains like classification and regression, evaluation metrics play a vital role in measuring model performance. Within medical diagnostics, metrics such as precision and recall are particularly significant. Precision reflects the proportion of correct positive predictions, meaning that when a model identifies cancer, it is likely to be correct—thereby reducing unnecessary alarm (Sokolova and Lapalme, 2009). Recall, on the other hand, focuses on identifying as many actual positive cases as possible, which is crucial in healthcare settings to avoid overlooking genuine cases of disease (Sokolova and Lapalme, 2009). Although accuracy is a common metric indicating the percentage of correctly predicted instances overall, it can be deceptive in scenarios with class imbalance, as it does not account for the nature of errors made (Tharwat, 2020).

To address such limitations, the F1-score is often employed. This metric, calculated as the harmonic mean of precision and recall, offers a more balanced view by considering both false positives and false negatives—an essential factor in medical imaging, where failing to detect malignant tumors can have serious consequences (Chicco and Jurman, 2020). Additionally, the area under the curve (AUC) of the receiver operating characteristic (ROC) curve is another key metric. It evaluates the model’s capacity to distinguish between positive (cancerous) and negative (non-cancerous) cases, with scores approaching 1 indicating high discriminative ability (Yang and Ying, 2022). Given the frequent imbalance in medical datasets, relying on a combination of these metrics provides a more reliable and holistic evaluation of model effectiveness (see Table 8).

The initial CNN architecture, developed for binary classification of histopathology images into benign and malignant categories, was evaluated using different filter and kernel size configurations. Table 9 presented in this section summarizes the performance metrics for each configuration. While the model achieved a maximum accuracy of 84%, this result was below the expected benchmark for binary classification tasks, where an accuracy of 95% or higher is generally required to maintain robust performance during the transition to multi-class classification.

Upon using the proposed DenseNet121-based binary classification, the model demonstrated outstanding performance in distinguishing between benign and malignant breast cancer histopathology images. The model was trained on out-of-sample biopsy images. As shown in Table 10, the model achieved a test accuracy of 98.63%, indicating a high level of reliability in classification. Precision, recall, and F1-score values were also notably high at 0.9888, 0.9867, and 0.9877, respectively, reflecting the model’s ability to correctly identify malignant cases while minimizing false positives and false negatives. The ROC-AUC score of 0.9863 further confirms the model’s strong discriminative capability, showing its robustness in distinguishing between the two classes across different classification thresholds. Figure 8 shows the confusion matrix for the classification, while Figure 9 is a graph showing the change in accuracy and loss with respect to epochs. Compared to previous models that relied on shallow convolutional architectures and dataset fragmentation based on magnification levels, the proposed model benefits from multi-scale feature extraction at different network depths, improving its ability to generalize across varying histopathological patterns. By utilizing a 128 × 128 input image resolution, the model strikes a balance between computational efficiency and feature preservation, allowing for optimal learning without excessive memory consumption. The use of feature fusion across different abstraction levels enhances the classification robustness, ensuring that both low-level structural features and high-level morphological variations contribute to the final decision-making process.

The performance of the multi-class classification models for benign and malignant breast cancer subtypes is summarized in Table 11. The model was trained on out-of-sample biopsy images. The model trained for benign classification achieved a test accuracy of 0.9382, whereas the malignant classification model obtained a test accuracy of 0.9158. Additionally, the malignant model demonstrated strong predictive capabilities, with a precision of 0.9151, recall of 0.9240, and an F1-score of 0.9182, while the benign model had a precision of 0.9541, recall of 0.9324 and an F1-score of 0.9415. These results highlight the effectiveness of using separate CNNs for each subtype, allowing for improved feature extraction and discrimination between fine-grained morphological variations. Figure 10 shows the confusion matrix for the benign model, while Figure 11 shows the same for the malignant model. Figure 12 shows the graph of the change in accuracy/loss with respect to the epochs for benign, and Figure 13 shows the same for malignant.

The results indicate that the model for benign subtypes outperformed the malignant classification model in terms of overall accuracy. However, the malignant classification model maintained strong recall and F1-score values, demonstrating its ability to correctly identify malignant subtypes while balancing precision. These findings support the hypothesis that training separate models for benign and malignant subtypes allows for more specialized learning, leading to improved classification performance. The integration of DenseNet121 as a feature extractor has further contributed to the model’s ability to capture multi-scale morphological patterns in histopathology images.

To evaluate the generalizability of the proposed classifiers, we performed 5-fold cross-validation on each of the three tasks: binary classification, benign subtype classification, and malignant subtype classification. The mean values, standard deviations, and 95% confidence intervals are presented in Table 14.

As shown in Table 14, the proposed classifiers maintained strong and stable performance across 5-fold cross-validation, with F1-scores of 0.9738 for binary classification, 0.9369 for benign subtype classification, and 0.8907 for malignant subtypes. The relatively tight confidence intervals and low standard deviations across all tasks suggest minimal sensitivity to training split variability, indicating good model generalization. Notably, the benign subtype classifier showed slightly higher performance consistency than the malignant subtype model, which exhibited greater variance in recall (±0.0265) and a broader 95% CI (0.8440–0.9176). This is likely due to the greater morphological diversity and class imbalance among malignant categories—particularly LC—where lower sample size and higher inter-class overlap may amplify prediction uncertainty.

The observed decline in recall and F1-scores compared to the single train-test split is expected, and it reflects the increased rigor of cross-validation, which exposes the model to harder-to-classify samples in more diverse folds. From a diagnostic perspective, the relatively stable precision across tasks implies that the model is conservative in positive predictions, minimizing false positives—an important trait in histopathological screening scenarios. However, the slightly larger variation in recall for rare or borderline subtypes warrants attention, as it may affect sensitivity to clinically ambiguous cases.

These findings reinforce the utility of k-fold cross-validation as a stress-testing tool, revealing not just average performance but also subtype-specific reliability under different sampling conditions. Future iterations of the model could incorporate stratified sampling by subtype or uncertainty-aware training strategies to further reduce variability in underrepresented classes.

The final model, built on DenseNet121 with multi-scale feature extraction, demonstrated the strongest performance among all evaluated architectures, achieving high accuracy, recall, and F1-scores in the binary classification of breast cancer biopsy images. By extracting and fusing features from three intermediate convolutional blocks, the model was able to learn both fine-grained cellular structures and broader tissue-level patterns relevant to histopathological classification. The incorporation of L2 normalization, dropout regularization, and batch normalization within each branch further contributed to its stability and generalization. This design effectively addressed limitations observed in earlier models, such as feature loss due to single-layer reliance or overfitting from insufficient regularization.

Compared to existing models evaluated on the same BreaKHis dataset, the proposed approach offers a substantial improvement. Araújo et al. (2017) employed a standard CNN and achieved an average accuracy of 83.3% across all magnification levels. Similarly, Vo et al. (2019) reported 86.3% accuracy using a handcrafted feature pipeline with SVMs, while Bayramoglu et al. (2016) applied a multi-scale CNN and reached approximately 88.0% accuracy. More recently, Munshi et al. (2024) proposed an ensemble CNN-SVM framework with XAI components, achieving 94.2% accuracy and an F1-score of 0.93. In contrast, our DenseNet121-based model consistently achieved >97% accuracy and an F1-score of 0.9738 in binary classification, while maintaining robust generalization across folds and magnifications. These improvements can be attributed to the deeper architectural depth, feature reuse enabled by dense connectivity, and the incorporation of multi-scale fusion, which allowed the model to capture both cellular and architectural histological patterns more effectively.

The model’s strong and consistent performance across both binary and subtype classification tasks positions it as a scalable and technically rigorous approach for digital histopathology. With additional validation on multi-institutional and heterogeneous datasets, this framework has the potential to contribute meaningfully to diagnostic support systems for breast cancer, especially in settings with limited expert pathologist availability.