Data Modality Determines Machine Learning and Deep Learning Success in Breast Cancer Prediction

A comprehensive comparative evaluation across differing data modalities has demonstrated that the diagnostic performance of artificial intelligence in breast cancer detection depends strictly on matching algorithmic architecture to input data types rather than assuming inherent model superiority. In an experimental investigation evaluating classical machine learning, deep learning, and transfer learning, researchers examined how algorithmic families behave when applied strictly to the data formats they were engineered to address. Utilizing the Wisconsin Diagnostic Breast Cancer structured tabular dataset alongside the Invasive Ductal Carcinoma histopathology image dataset, the study revealed that classical linear and margin-based classifiers attained up to 98.24 percent accuracy on preprocessed numerical features. Conversely, deep convolutional neural networks were indispensable for analyzing raw microscopic tissue slides, where the transfer learning model DenseNet121 achieved 91.2 percent accuracy. The findings systematically challenge common literature trends where deep learning models are inappropriately forced onto compact tabular registries or classical algorithms are tasked with interpreting flattened image arrays without spatial context.

The study was motivated by persistent methodological discrepancies in published computational oncology literature, where algorithms of disparate paradigms are frequently pitted against one another under inequitable conditions. Classical machine learning frameworks excel when operating on compact numerical tables characterized by explicit measurements, whereas deep neural networks require high-dimensional unstructured arrays such as digitized pathological slides where contextual textures and biological boundaries dictate diagnosis. To resolve these comparative imbalances, the investigators established an equitable, dual-track experimental framework executed on a Windows 10 computing workstation equipped with an Intel Core i5 processor and implemented in Python 3.9 utilizing the pandas data science environment. Rather than applying deep learning directly to tabular records where high parameter counts lead to severe overfitting, or forcing tabular classifiers to read raw images, each paradigm was evaluated on its native modality before a principled cross-paradigm assessment of diagnostic accuracy, computational efficiency, and biological interpretability was established.

For the structured clinical analysis, the researchers employed the benchmark Wisconsin Diagnostic Breast Cancer dataset, which comprises 569 patient samples derived from fine needle aspiration biopsy procedures. Each record contains 30 continuous numerical attributes capturing nuclear morphological properties, including perimeter, texture, area, radius, compactness, smoothness, concavity, concave points, symmetry, and fractal dimension, categorized across mean, standard error, and extreme values. The biological distribution reflects 357 benign cases representing 62.7 percent of the cohort and 212 malignant cases representing 37.3 percent, with zero missing values detected during preliminary pipeline verification. To guarantee mathematical stability across features exhibiting divergent scales—such as cell area extending from 143 to 2501 versus fractal dimensions ranging between 0.05 and 0.20—the tabular pipeline implemented an initial log1p transformation to compress right-skewed profiles into approximately normal distributions prior to data partitioning.

To rigorously eliminate information leakage, the tabular data was split into an 80 percent training partition containing 455 samples and a 20 percent testing partition containing 114 samples using stratified sampling to preserve the 37.3 to 62.7 target balance. Crucially, feature standardization via StandardScaler was fitted strictly on training records alone before transforming held-out test records, preventing indirect leakage of test distribution statistics into model training. Within this leakage-free protocol, five classical algorithms were tested: Logistic Regression, Support Vector Machine with a radial basis function kernel, Random Forest ensembles with 200 estimators, K-Nearest Neighbors configured with seven neighbors to prevent tie voting, and Decision Tree classifiers. Hyperparameter optimization through randomized cross-validation revealed that Logistic Regression and Support Vector Machines were exceptionally stable, both achieving 97.37 percent accuracy and 97.90 percent F1-scores on the test set, while hyperparameter-tuned Logistic Regression reached 98.24 percent accuracy under the primary split and 98.43 percent under an alternative 70:30 validation split.

To confirm whether deep neural networks could offer any marginal utility on structured clinical measurements, the authors constructed four tailored architectures for the tabular dataset: a Multi-Layer Perceptron, a one-dimensional Convolutional Neural Network, a residual ResNet-MLP, and an Attention-MLP utilizing dynamic attribute weighting. The Multi-Layer Perceptron yielded the highest deep learning accuracy at 94.74 percent with an F1-score of 95.71 percent and an area under the receiver operating characteristic curve of 0.9944, followed closely by ResNet-MLP at 93.86 percent. In contrast, the one-dimensional convolutional model performed poorly, recording only 85.96 percent accuracy and an 88.06 percent F1-score due to the absence of meaningful spatial adjacency between tabular columns. Crucially, none of the neural networks matched the 98.24 percent accuracy or 97.90 percent F1-score attained by classical Logistic Regression. A paired t-test between Logistic Regression and the Multi-Layer Perceptron across cross-validation folds yielded a t-value of negative 0.492 and a p-value of 0.648, confirming that while the simpler linear model achieved higher consistency, the margin remains a statistically non-significant trend.

For the visual histopathology track, the investigation examined the Invasive Ductal Carcinoma Regular Breast Histopathology dataset, extracted from 279 digitized biopsy whole-slide images. The complete repository encompasses 277,524 microscopic patches sized at 50 by 50 pixels, displaying an uncurated natural imbalance of approximately 78 percent benign tissue and 22 percent malignant ductal carcinoma. To construct a computationally tractable training set within physical GPU memory boundaries, the researchers assembled a balanced subset of 6,000 image patches, consisting of exactly 3,000 benign and 3,000 malignant samples. This corpus was partitioned via an 80:10:10 protocol into 4,800 training images, 600 validation images, and 600 unseen test images. Each image patch was scaled to an input resolution of 224 by 224 pixels and subjected to biological data augmentation, including random rotations, zooming, and vertical and horizontal reflections, which appropriately mirror the unoriented nature of histological slides.

A critical architectural contribution of the deep learning pipeline involved identifying and rectifying a widespread data processing error that frequently corrupts published computer vision benchmarks. In many public implementations, pixel normalization by a factor of 255 is performed simultaneously within the image loading script and inside the neural network model layers, causing double normalization that collapses pixel intensities toward zero and yields near-random classification accuracy around 51 percent. By restricting floating-point pixel rescaling strictly to a single internal model layer while delivering raw pixel values from the data loader, the pipeline maintained genuine morphological signal integrity. The researchers subsequently evaluated three convolutional architectures on the balanced IDC cohort: a customized baseline three-block Convolutional Neural Network trained entirely from scratch with the Adam optimizer at a learning rate of 0.0001, a pre-trained EfficientNetB0 backbone, and a pre-trained DenseNet121 network utilizing a two-stage transfer learning strategy.

Experimental evaluation of the image models demonstrated that DenseNet121 was the superior architecture for histopathological tissue evaluation. When executed with an initial frozen feature extraction phase followed by unfreezing its final 50 layers with a refined learning rate, DenseNet121 achieved an overall accuracy of 91.2 percent, a precision of 91.27 percent, a recall of 91.17 percent, an F1-score of 91.16 percent, and an area under the receiver operating characteristic curve of 0.962. It also recorded the lowest false-negative count, misclassifying only 19 malignant patches. The baseline Convolutional Neural Network trained from scratch delivered stable, moderate convergence, reaching 84.5 percent accuracy and an area under the curve of 0.9349, though it produced 71 false negatives. Conversely, EfficientNetB0 failed entirely to adapt to the histopathological domain, hovering at near-random accuracy of 50.00 percent and an area under the curve of 0.5000 in both frozen and fine-tuned configurations, proving that ImageNet pre-training cannot universally overcome extreme architectural mismatches in medical textures without domain alignment.

To explore the limits of machine learning on imagery, the investigators transformed the histological patches into 912-dimensional numerical vectors using combined Histogram of Oriented Gradients and color descriptors, subsequently applying Principal Component Analysis to reduce the feature space to 292 components capturing 95 percent of variance. When classical models were deployed on these flattened representations, the top-performing Support Vector Machine reached only 82.17 percent accuracy and an F1-score of 82.15 percent, trailing DenseNet121 by roughly nine percentage points. When generalizability was checked by scaling deep models onto the entire imbalanced collection of 277,524 patches with class weighting, the baseline convolutional model improved its F1-score from 0.844 to 0.889, while DenseNet121 remained robust with an F1-score of 0.876 and an area under the curve of 0.950, confirming the structural necessity of spatial convolutional operations over flattened vectors.

Transparency analysis using Explainable Artificial Intelligence tools further illuminated internal decision dynamics while highlighting notable interpretability constraints. In the tabular track, SHAP explanations showed that worst texture, radius error, worst radius, and worst area exerted the strongest influence on Logistic Regression predictions. On the histopathology images, Grad-CAM, SHAP, and LIME identified cellular abnormalities and glandular clusters, but quantitative faithfulness evaluations via Area Over the Perturbation Curve revealed weak grounding, alongside minimal pairwise agreement across explanation methods. Computational profiling established that while Logistic Regression completed inference in 0.12 seconds, DenseNet121 required 540.4 seconds of training and 5.70 billion floating-point operations. The authors explicitly concluded that despite robust classification metrics, their dual-track framework serves strictly as a clinical decision-support reference and cannot replace certified pathologists without extensive multi-center clinical validation, scanner calibration, and prospective regulatory evaluation.

Subject of Research: Comparative evaluation of machine learning and deep learning architectures across structured tabular and histopathological image data modalities for breast cancer classification

Article Title: Explainable breast cancer classification using machine learning deep learning and transfer learning across multiple data modalities

Article References: Tiwari, B., Mathur, P., & Chaurasia, B. K. (2026). Explainable breast cancer classification using machine learning deep learning and transfer learning across multiple data modalities. Discover Computing, 29(1), Article 888. https://doi.org/10.1007/s10791-026-10782-x

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10782-x

Keywords: Breast Cancer Classification, Machine Learning, Deep Learning, Transfer Learning, Wisconsin Diagnostic Breast Cancer, Invasive Ductal Carcinoma, DenseNet121, Logistic Regression, Support Vector Machine, Histopathology Imaging, Explainable Artificial Intelligence, SHAP, LIME, Grad-CAM, Data Modality, Principal Component Analysis, EfficientNetB0, Diagnostic Modeling