[Paper Review] Data Augmentation for Detection of Architectural Distortion in Digital Mammography using Deep Learning Approach

Arthur C. Costa, Helder C. R. Oliveira+4 more|arXiv (Cornell University)|Jul 6, 2018

AI in cancer detection6 references12 citations

TL;DR

This study proposes a data augmentation strategy to train a Convolutional Neural Network (CNN) for detecting architectural distortion (AD) in digital mammography, a subtle early sign of breast cancer. Despite a limited dataset of 300 clinical images, data augmentation increased training samples to 21,600, enabling the CNN to achieve an AUC of 0.74 on a realistic test set, demonstrating feasibility for early cancer detection using deep learning with constrained data.

ABSTRACT

Early detection of breast cancer can increase treatment efficiency. Architectural Distortion (AD) is a very subtle contraction of the breast tissue and may represent the earliest sign of cancer. Since it is very likely to be unnoticed by radiologists, several approaches have been proposed over the years but none using deep learning techniques. To train a Convolutional Neural Network (CNN), which is a deep neural architecture, is necessary a huge amount of data. To overcome this problem, this paper proposes a data augmentation approach applied to clinical image dataset to properly train a CNN. Results using receiver operating characteristic analysis showed that with a very limited dataset we could train a CNN to detect AD in digital mammography with area under the curve (AUC = 0.74).

Motivation & Objective

• To address the challenge of limited clinical data for training deep learning models in mammography.
• To improve detection of architectural distortion (AD), a subtle and early sign of breast cancer often missed by radiologists.
• To develop and evaluate a data augmentation strategy that expands small clinical datasets for effective CNN training.
• To assess the performance of the trained CNN on a realistic, clinically relevant test scenario.

Proposed method

• The study used a dataset of 300 clinical mammograms, including 200 from the DDSM dataset and 100 acquired under institutional review board approval.
• Regions of Interest (ROIs) were manually cropped based on radiologist-annotated AD locations and normal tissue, resulting in 600 ROIs.
• Data augmentation applied vertical/horizontal flips, 90°, 180°, and 270° rotations, and Gaussian noise (variance 0.02, 0.04, 0.06), increasing the dataset to 21,600 samples with balanced classes.
• A CNN was implemented using TensorFlow with 5x5 filters, max pooling, and a 4x4 fully connected layer, trained on a 70%-15%-15% split with z-score normalization.
• The model was evaluated on a test set of 9 new exams, where ~3,000 ROIs per exam were extracted and classified as AD or normal based on center coordinates.

Experimental results

Research questions

• RQ1 Can data augmentation effectively compensate for limited clinical data in training a CNN for architectural distortion detection in mammography?
• RQ2 How well does a CNN trained on augmented data generalize to new, unseen mammographic exams in a realistic clinical setting?
• RQ3 What is the diagnostic performance of the proposed method in terms of AUC and accuracy on a realistic test scenario?

Key findings

• The CNN achieved an AUC of 0.74 on a realistic test set of nine new mammographic exams, demonstrating practical diagnostic potential despite limited training data.
• On the training and validation split, the model reached 99.4% accuracy and an AUC of 0.99, indicating strong internal learning.
• The data augmentation strategy successfully expanded the initial 600 ROIs to 21,600 samples, enabling effective CNN training with balanced classes.
• The model's performance on the realistic test set (AUC=0.74, accuracy=86.1%) confirms its potential for clinical deployment in early breast cancer detection.