Breast cancer remains one of the leading causes of death for women globally. Early detection through screening is critical, and mammography is the primary screening tool - though accurate interpretation of mammogram images is demanding and subject to human error.
The two most important early signs of breast cancer in mammograms are masses (tissue growths that may be benign or malignant) and microcalcifications (MCs) (tiny calcium deposits that can indicate early cancer). Malignant masses tend to have irregular outlines and appear denser than surrounding tissue, but distinguishing them from benign masses is not always straightforward.
Computer-aided detection (CAD) systems were developed to assist radiologists by automatically flagging suspicious regions. Traditional CAD approaches relied on hand-crafted feature extraction methods. More recently, deep learning - particularly deep convolutional neural networks (DCNNs) - has demonstrated superior ability to learn complex features automatically from mammogram images.
This study proposes a new CAD system that uses a fine-tuned AlexNet deep CNN for feature extraction, combined with a support vector machine (SVM) classifier at the final stage. The key insight is that while DCNNs excel at learning visual features, replacing the network's last classification layer with a dedicated SVM classifier can yield better discrimination between benign and malignant masses.
The system also tests two different methods for isolating the tumor region before feature extraction: manual cropping using circular contours, and an automated threshold and region-based segmentation approach. Both are evaluated on standard public mammography datasets.
The study demonstrates that the hybrid DCNN-SVM architecture achieves an AUC of 0.94 (94%) on the CBIS-DDSM dataset - the highest reported at the time of publication using AlexNet-based architectures - showing that combining the feature learning strengths of deep networks with the classification power of SVMs produces meaningful improvements.
The first step is image enhancement using Contrast-Limited Adaptive Histogram Equalization (CLAHE), which improves local contrast in mammogram images to make subtle abnormalities more visible while limiting the over-amplification of noise that can occur with standard contrast enhancement methods.
The second step is image segmentation - extracting the region of interest (ROI) containing the tumor from the broader mammogram. Two methods were tested: (1) manual extraction using predefined circular contours around labeled tumors; and (2) automatic extraction by applying a threshold to detect the red-labeled contour surrounding each tumor in the DDSM dataset, then selecting the largest enclosed region.
For feature extraction, the study uses AlexNet, a landmark DCNN architecture with five convolutional layers, three pooling layers, and two fully connected layers containing approximately 60 million parameters. AlexNet was originally trained on ImageNet for 1,000-class general image classification. Here it is adapted using transfer learning - its final layer is replaced to classify only two classes: benign and malignant.
To address the limited size of medical imaging datasets, data augmentation was applied by rotating each image by 0, 90, 180, and 270 degrees, effectively quadrupling the number of training samples. The model was trained with 70% of data and tested on the remaining 30%, using 5-fold cross-validation for validation.
A support vector machine (SVM) is a supervised machine learning algorithm designed to find the optimal boundary - called a hyperplane - that separates two classes of data with the maximum possible margin between them. The data points closest to this boundary are called support vectors, which define the decision boundary.
SVMs can operate in high-dimensional feature spaces, making them well-suited to process the rich feature representations extracted by deep neural networks. The features output by AlexNet's final fully connected layer are fed directly into the SVM for the final benign vs. malignant classification decision.
Several SVM kernel functions were compared - including linear, quadratic, cubic, and Gaussian variants. The kernel function controls how the SVM measures similarity between data points in feature space. Results showed that linear and medium Gaussian kernels achieved the best classification performance depending on the dataset used.
On the DDSM dataset using manually cropped ROIs, AlexNet alone achieved 71.01% accuracy. Adding the SVM classifier improved this to 79% accuracy with AUC of 0.88 using a linear kernel. The threshold-based automated segmentation performed slightly better, yielding 80.5% accuracy with AUC 0.88, also with a linear kernel.
The best results came from the CBIS-DDSM dataset, which contained images that were already professionally segmented. AlexNet alone reached 73.6% accuracy. The DCNN-SVM combination achieved 87.2% accuracy with AUC of 0.94 using a medium Gaussian kernel, along with sensitivity of 86.2%, specificity of 87.7%, precision of 88%, and F1 score of 87.1%.
Across all experiments, adding the SVM stage consistently improved accuracy by 6-14 percentage points over the DCNN alone. The better performance on CBIS-DDSM compared to DDSM is attributed to the higher-quality pre-segmented ROI data in that dataset, underscoring how segmentation quality affects downstream classification performance.
Compared to other AlexNet-based systems, this study achieved the highest AUC reported for the same architecture and similar conditions. Competing approaches achieved AUCs of 0.81 (Huynh and Giger) and 0.83 (Jiang), while the proposed DCNN-SVM system reached 0.94 on CBIS-DDSM.
When compared across all DCNN architectures and datasets, the proposed system also ranked highest in AUC among reported results, including approaches using VGG networks, GoogLeNet, and dedicated CNN architectures trained on different mammography datasets.
The results validate the core hypothesis: that using a deep neural network as a feature extractor and a separate SVM as the classifier is more effective than using the neural network's own built-in classification layer alone. The combination leverages the strengths of both methods.
The study demonstrates that combining a fine-tuned deep CNN (AlexNet) for feature extraction with an SVM for final classification produces a strong breast cancer CAD system, achieving an AUC of 0.94 on the CBIS-DDSM dataset - the highest reported using the same architecture at time of publication.
The automated region-based segmentation method performed comparably to manual ROI cropping, suggesting that fully automated end-to-end pipelines - eliminating the need for manual annotation - are feasible without significant accuracy loss.
Future work will extend the system to deeper architectures such as VGG and ResNet, and to detecting other breast abnormalities beyond masses, including microcalcifications. These improvements aim to create a more comprehensive CAD tool capable of assisting radiologists across the full range of mammographic findings.