Breast cancer is the second leading cause of cancer death among women worldwide, accounting for 11.7% of all female cancer deaths. Early and accurate detection is critical for improving treatment outcomes and survival rates.
Medical imaging techniques such as mammography, ultrasound, MRI, PET, and CT are used to detect breast cancer. Mammography is the gold standard for screening, but interpreting the images accurately at scale remains challenging and labor-intensive for radiologists.
Convolutional neural networks (CNNs) have shown strong potential for automating and improving the accuracy of tumor detection and classification in medical images. However, deep learning models require careful tuning of their configuration settings, known as hyperparameters, to perform optimally - and doing this manually is time-consuming and imprecise.
This study proposes a hybrid framework called EACO-ResNet101, which combines a powerful pre-trained deep learning architecture (ResNet101) with an improved optimization algorithm to automatically find the best configuration settings for the model.
The optimization component is the Enhanced Ant Colony Optimization (EACO) algorithm, which improves on the standard Ant Colony Optimization technique by incorporating opposition-based learning (OBL). This combination allows the algorithm to more thoroughly explore possible model configurations and avoid getting stuck in suboptimal solutions.
The framework was tested on two well-established mammography datasets - CBIS-DDSM and MIAS - and achieved classification accuracy above 98% on both, outperforming all comparison methods tested in the study.
Ant Colony Optimization (ACO) is a nature-inspired algorithm that mimics how ant colonies find the shortest path to food. Ants deposit chemical trails called pheromones as they walk; shorter paths accumulate stronger trails over time, causing the whole colony to converge on the optimal route.
In computing, ACO is used to search complex spaces for optimal solutions - in this case, the best combination of neural network hyperparameters. However, standard ACO can converge prematurely and miss better solutions by getting stuck exploring too narrow a region of the search space.
Opposition-based learning (OBL) addresses this by simultaneously evaluating a solution and its mathematical opposite, selecting whichever performs better. When integrated into the ACO framework as EACO, this technique improves diversity in the search process, helping the algorithm escape local optima and find better hyperparameter configurations than standard ACO alone.
ResNet101 is a deep CNN with 101 layers, pre-trained on the large ImageNet image database. Transfer learning allows the model to reuse the visual features learned from millions of general images and adapt them to the specific task of classifying mammograms, avoiding the need to train from scratch on limited medical imaging data.
The proposed framework operates in four stages: (1) pre-processing and augmentation, where images are resized to 224x224 pixels and artificially varied to expand the training set; (2) hyperparameter optimization using EACO; (3) model training on the target datasets; and (4) performance evaluation using standard metrics.
The EACO algorithm optimizes eight hyperparameters: learning rate, batch size, three dropout rates, and the number of neurons in three dense layers. These parameters collectively control how the model learns, how aggressively it generalizes, and how quickly it converges during training.
Data augmentation techniques including shearing, zooming, rotation, and flipping were applied to the training sets to artificially increase dataset size and reduce overfitting - a common problem where models learn training data too precisely and fail to generalize to new images.
On the CBIS-DDSM dataset (5,482 mammography images), EACO-ResNet101 achieved 98.63% accuracy, 98.76% sensitivity, and 98.89% specificity. This represents an 8.42% improvement in accuracy over the same ResNet101 model configured without hyperparameter optimization (90.21%).
On the MIAS dataset (344 images, expanded with augmentation to 1,376), the model achieved 99.15% accuracy, 97.86% sensitivity, and 98.88% specificity - an 11.48% accuracy improvement over baseline ResNet101 (87.67%). These results demonstrate that automatic hyperparameter tuning produces substantial gains regardless of dataset size.
Sensitivity measures how well the model identifies true cancer cases, while specificity measures how well it avoids false alarms in healthy tissue. Both scores above 97% indicate the model performs reliably across both types of classification errors - crucial for clinical applications where both missed cancers and unnecessary follow-up procedures carry real costs.
EACO was compared against four other well-known metaheuristic optimization algorithms: Gravitational Search Algorithm (GSA), Harris Hawks Optimization (HHO), Particle Swarm Optimization (PSO), and standard ACO. All were used to optimize the same ResNet101 architecture under identical conditions.
On CBIS-DDSM, competing methods achieved accuracies ranging from 94.7% (HHO) to 96% (ACO), compared to EACO's 98.63%. On MIAS, competing methods ranged from 93.5% (PSO) to 95.1% (ACO), compared to EACO's 99.15%. EACO outperformed all alternatives on every metric measured.
The results suggest that integrating opposition-based learning into the ant colony framework provides a meaningful advantage in hyperparameter search quality. The improvement over standard ACO in particular demonstrates that the OBL enhancement is responsible for a significant portion of the performance gains.
The primary clinical goal of this research is to assist radiologists in rapidly identifying anomalies in mammography images. Given the high volume of screening mammograms and the cognitive demands of image interpretation, AI tools that can reliably pre-screen or flag suspicious regions could substantially reduce radiologist workload and diagnostic error rates.
A classification accuracy above 98% across both benign and malignant categories - with balanced sensitivity and specificity - suggests the model could serve as a reliable second opinion or screening aid in clinical practice, particularly in settings where access to specialized radiologists is limited.
The authors acknowledge that the model was tested on curated benchmark datasets and that future work should include larger, more diverse real-world datasets, as well as integration of additional pre-trained architectures such as DenseNet and Inception to further improve reliability and generalizability.
This study demonstrates that combining a powerful pre-trained CNN architecture with an intelligently designed optimization algorithm can significantly improve breast cancer classification accuracy from mammography images, achieving above 98% on two standard datasets.
The key innovation is the EACO algorithm, which automates the challenging process of hyperparameter tuning by borrowing from the collective intelligence of ant colonies and enhancing it with opposition-based search. This removes a significant barrier to deploying high-performance deep learning models in medical imaging applications.
Future work will extend the model to larger datasets, incorporate additional pre-trained architectures, and optimize the balance between data augmentation and dataset diversity. The authors aim to further increase classification accuracy and reliability, moving the technology closer to clinical deployment for routine breast cancer screening.