Source-linked AI summary
Breast Mass Classification from Mammograms using Deep Convolutional Neural Networks
Daniel Lévy, Arzav Jain
TL;DR
Mammography screening is manual and can miss or misdiagnose breast masses because appearance varies and image signal is limited. The paper trains end-to-end CNNs to classify pre-segmented masses using transfer learning, preprocessing, and augmentation. Its best model achieves 0.934 recall at 0.924 precision, exceeding the cited radiologist recall range, while saliency maps provide interpretability.
Problem
Manual mammography interpretation is vulnerable to human error from variable mass appearance, low signal-to-noise ratio, and workload differences, while prior breast-mass methods were not fully end-to-end.
Method
The paper trains CNNs to classify pre-segmented breast masses directly from mammogram pixels, using transfer learning, preprocessing, data augmentation, and saliency-map visualization.
Results
0.934 recall at 0.924 precision is achieved by the best model, compared with radiologist recall between 0.745 and 0.923; GoogLeNet outperforms the other models.
Takeaways & Limitations
The results support end-to-end CNN classification as a promising approach for breast-mass diagnosis, with interpretability intended to facilitate practical adoption.
Abstract
from arXiv · showhide
Mammography is the most widely used method to screen breast cancer. Because of its mostly manual nature, variability in mass appearance, and low signal-to-noise ratio, a significant number of breast masses are missed or misdiagnosed. In this work, we present how Convolutional Neural Networks can be used to directly classify pre-segmented breast masses in mammograms as benign or malignant, using a combination of transfer learning, careful pre-processing and data augmentation to overcome limited training data. We achieve state-of-the-art results on the DDSM dataset, surpassing human performance, and show interpretability of our model.
1 Introduction
Mammography is widely used for breast-cancer screening, but manual interpretation is vulnerable to variable mass appearance, low signal-to-noise ratio, human error, and workload-related differences. The paper addresses these challenges by directly classifying pre-segmented masses with end-to-end CNNs using transfer learning, preprocessing, and augmentation.
- Mammography supports preventive breast-cancer detection but remains manual and prone to missed malignancies or unnecessary biopsies.Performance is also correlated with radiologist expertise and workload.
- End-to-end CNNs directly classify pre-segmented mammographic masses as benign or malignant.The paper addresses the scarcity of available data and interpretability concerns associated with prior approaches.
- Transfer learning, careful preprocessing, and data augmentation produce state-of-the-art results while supporting interpretability analysis.The paper also analyzes the effects of these modeling choices.
2 Related Work
Prior medical-image work used CNNs and multi-stage or heavily engineered pipelines for mammography. This paper positions its contribution as direct CNN classification of pre-detected breast masses.
- CNNs have achieved strong results across classification, object detection, and segmentation in computer vision and medical imaging.Medical applications include chest pathology identification, lymph-node detection, and interstitial lung-disease classification.
- Mammography studies commonly use multi-stage pipelines involving detection, engineered features, or aggregation across breast views.Examples include R-CNNs with random forests, semantic and textural features, and full-mammogram view aggregation.
- The authors identify their approach as the first to directly classify pre-detected breast masses using CNN architectures.This contrasts with prior work requiring extensive domain-knowledge preprocessing or other intermediate stages.
3 Dataset
The study uses the public DDSM mammography dataset, retaining images containing masses and splitting them by patient into balanced validation and test sets. Figure 1 illustrates benign and malignant mass examples.
- DDSM contains approximately 2500 studies with mediolateral oblique and craniocaudal breast views, grayscale images, and masks for pre-segmented masses.The dataset is collaboratively maintained at the University of South Florida.
- The mass-containing subset includes 1820 images from 997 patients.Only mammograms containing masses were considered.
- Patient-level random splits allocate 80% to training, 10% to validation, and 10% to testing, with balanced validation and test sets.
- Figure 1 shows benign masses on the left and malignant masses on the right.
4 Methods
The method compares three CNN architectures and varies transfer learning, context size, and augmentation to classify pre-segmented masses. Context is supplied through fixed or proportional padding, while augmentation expands the effective training set.
- CNN architectures: The study evaluates a shallow baseline CNN, AlexNet, and GoogLeNet for binary breast-mass classification.AlexNet and GoogLeNet replace their final fully connected layers with two-class outputs, and GoogLeNet’s auxiliary classifiers are removed.
- CNN architectures: The baseline CNN uses three convolutional blocks followed by fully connected layers and a two-class softmax output.It accepts 224 × 224 × 3 inputs and uses batch normalization, ReLU activations, Xavier initialization, and Adam.
- Transfer learning: ImageNet pretraining initializes convolutional layers before fine-tuning on mammography images, with layer-specific learning-rate multipliers.Fully connected layers are randomly initialized, while GoogLeNet receives more aggressive learning in its final layer.
- Mass context: Small Context adds 50 pixels of fixed padding, whereas Large Context extracts a region twice the mass bounding-box size.These alternatives provide context around the mass using fixed or proportional padding.
- Data augmentation: Augmentation applies rotations, crops, and mirroring, increasing the effective training-set size by a factor of 25.Five random rotations and five random crops per rotation are generated offline, with random mirroring at training time.
5 Results
The experiments show that transfer learning, proportional mass context, and augmentation improve the CNN pipeline, with GoogLeNet achieving the strongest final performance. The best model reaches high recall and precision, while saliency maps indicate attention to diagnostically relevant mass edges and surrounding context.
- Transfer learning: Fine-tuned AlexNet significantly outperforms the baseline, confirming the usefulness of ImageNet pre-training for limited mammography data.
- Mass context: Proportional padding around the mass contains greater classification signal than fixed padding and is used in subsequent experiments.
- Data augmentation: Data augmentation regularizes the network and helps address data scarcity by increasing the effective training data.The augmentation compares training on unaugmented and augmented datasets across iterations.
- Final performance: GoogLeNet outperforms the other tested architectures and is less prone to overfitting, with approximately 5 million parameters versus 100 million for AlexNet.
- Final performance: 0.934 recall at 0.924 precision is achieved by the best model, exceeding the reported radiologist recall range of 0.745–0.923.
- Interpretability: Saliency maps show AlexNet and GoogLeNet attending to mass edges and contextual regions in five validation images.Brighter regions indicate larger image gradients and higher contribution to predictions.
6 Conclusion
The work develops an end-to-end deep learning model for classifying pre-detected mammographic masses and combines preprocessing, augmentation, transfer learning, and interpretability methods. It reports state-of-the-art performance, exceeding trained radiologists, while interpretability supports more comfortable real-world adoption.
- The proposed model classifies pre-detected breast masses from mammograms end-to-end as benign or malignant.
- Careful pre-processing, data augmentation, and transfer learning address the data bottleneck in medical computer vision.
- The approach provides interpretability for network predictions through saliency-map visualization methods.
- The approach obtains state-of-the-art results and outperforms trained radiologists.
- Interpretability enables more comfortable adoption in real-world settings.