Related Experiment Video
Updated: Jul 29, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.0K
A divide and conquer approach to maximise deep learning mammography classification accuracies
Adam Jaamour1, Craig Myles1, Ashay Patel1
1School of Computer Science, University of St Andrews, St Andrews, Fife, United Kingdom.
Plos One
|May 26, 2023
Summary
Deep learning models can improve mammography accuracy for breast cancer detection. Combining transfer learning with specific data handling techniques achieved a 5.6% accuracy gain over the baseline model.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer is a leading cause of death in the UK, with mammography being crucial for early detection.
- Inaccurate mammography diagnoses can lead to adverse patient outcomes, necessitating improved diagnostic systems.
- Deep learning offers potential for automated breast cancer detection in mammograms, reducing interpretation errors.
Purpose of the Study:
- To explore various deep learning techniques for enhancing breast cancer detection in mammograms.
- To analyze the impact of different deep learning strategies on model performance and efficiency using a divide and conquer approach.
- To provide practitioners with data-driven insights for selecting optimal deep learning methods for mammography classification tasks.
Main Methods:
- Utilized Convolution Neural Networks (CNNs) within a deep learning pipeline.
- Employed a divide and conquer strategy to evaluate diverse techniques: network architectures (VGG19, ResNet50, InceptionV3, DenseNet121, MobileNetV2), class weights, input sizes, image preprocessing, transfer learning, dropout rates, and mammogram projection types.
- Applied transfer learning with pre-trained ImageNet weights to MobileNetV2, incorporating mini-MIAS dataset weights and class imbalance mitigation strategies on the CBIS-DDSM dataset.
Main Results:
- A combination of transfer learning (ImageNet weights to MobileNetV2), mini-MIAS pre-trained weights, class weights for imbalance, and splitting data by masses/calcifications yielded a 5.6% accuracy increase over the VGG19 baseline.
- Larger image sizes alone did not improve accuracy; effectiveness was contingent on image preprocessing techniques like Gaussian filtering, histogram equalization, and cropping.
- The divide and conquer approach identified specific deep learning techniques that enhance accuracy for mammography classification.
Conclusions:
- Optimized deep learning strategies, particularly transfer learning and tailored data handling, significantly improve breast cancer detection accuracy in mammograms.
- The study provides a practical framework for selecting effective deep learning models, reducing the need for extensive experimentation by clinicians.
- Further research can build upon these findings to develop more robust and accurate AI-assisted mammography systems, ultimately benefiting patient care.
