Source-linked AI summary
Detecting and classifying lesions in mammograms with Deep Learning
Dezső Ribli, Anna Horváth, Zsuzsa Unger, Péter Pollner, István Csabai
TL;DR
Screening mammography remains error-prone, and existing CAD has produced controversial benefits. This paper uses Faster R-CNN to detect and classify breast lesions, achieving strong performance on INbreast and second place in the Digital Mammography DREAM Challenge.
Problem
Screening mammography remains error-prone, while existing CAD systems show controversial benefits and need improvement.
Method
The study trains a Faster R-CNN CAD system on annotated mammograms to automatically detect, localize, and classify malignant or benign lesions.
Results
AUC = 0.85 earned second place in the Digital Mammography DREAM Challenge, while INbreast detection reached 90% malignant-lesion sensitivity with 0.3 false-positive marks per image.
Takeaways & Limitations
The proposed lesion detector could serve as a perception-enhancement tool to help radiologists detect more cancers.
Takeaways & Limitations
Detection performance could only be evaluated on the small INbreast dataset because publicly available pixel-level annotated datasets are limited.
Abstract
from arXiv · showhide
In the last two decades Computer Aided Diagnostics (CAD) systems were developed to help radiologists analyze screening mammograms. The benefits of current CAD technologies appear to be contradictory and they should be improved to be ultimately considered useful. Since 2012 deep convolutional neural networks (CNN) have been a tremendous success in image recognition, reaching human performance. These methods have greatly surpassed the traditional approaches, which are similar to currently used CAD solutions. Deep CNN-s have the potential to revolutionize medical image analysis. We propose a CAD system based on one of the most successful object detection frameworks, Faster R-CNN. The system detects and classifies malignant or benign lesions on a mammogram without any human intervention. The proposed method sets the state of the art classification performance on the public INbreast database, AUC = 0.95 . The approach described here has achieved the 2nd place in the Digital Mammography DREAM Challenge with AUC = 0.85 . When used as a detector, the system reaches high sensitivity with very few false positive marks per image on the INbreast dataset. Source code, the trained model and an OsiriX plugin are availaible online at https://github.com/riblidezso/frcnn_cad .
Introduction
Screening mammography reduces breast cancer mortality but remains vulnerable to missed cancers and reader errors. Because existing CAD has controversial clinical benefits, deep CNNs motivate improved mammogram analysis.
- Screening mammography was shown to reduce breast cancer mortality by 38–48% among participants.
- 20-30% of diagnosed cancers could be found retrospectively on previous negative screening exams by blinded reviewers.Missed cancers persist despite modern full field digital mammography.
- CAD programs mark suspicious mammographic regions for radiologist review, but their clinical benefits are controversial.Currently used CAD technologies were concluded in multiple studies not to improve radiologists’ everyday-practice performance, indicating a need for improvement.
- Since 2012, deep convolutional neural networks have significantly outperformed traditional computer-vision methods and reached or surpassed human performance in image classification and object detection.Current CAD approaches rely on hand-crafted features followed by machine-learning classification.
- The Digital Mammography DREAM Challenge used 86000 exams with breast-level binary labels indicating cancer diagnosis within the next 12 months.Each breast side was treated as a separate case, without pixel-level annotation.
1 Material and methods
The study trained a Faster R-CNN lesion detector on annotated DDSM and Semmelweis University mammograms and evaluated it on INbreast. The model used VGG16, joint optimization, image preprocessing, and augmentation tailored to mammography.
- Datasets: Training used pixel-level annotated mammograms from DDSM and Semmelweis University, while testing used the public INbreast dataset.DDSM was used only for training; the Semmelweis dataset supported improvement during the second stage of the Digital Mammography DREAM Challenge.
- Datasets: 115 INbreast FFDM cases were adapted by ignoring benign annotations, converting malignant lesions to bounding boxes, and excluding 8 exams with confounding findings or ambiguous outcomes.The INbreast cases had pixel-level annotations and histological proof for cancers.
- Model architecture: The model used Faster R-CNN with a VGG16 base CNN to detect benign or malignant lesions and output lesion bounding boxes with class-confidence scores.Image-level scores were derived from the maximum malignant-lesion score, averaged across multiple images of the same breast.
- Training: Training jointly optimized object detection and classification using backpropagation and stochastic gradient descent with weight decay.The initial model was pretrained on 1.2 million ImageNet images.
- Preprocessing and augmentation: Mammograms were isotropically rescaled to fit GPU memory, and vertical and horizontal flipping augmented the training data.The IoU threshold for foreground proposals was relaxed from 0.7 to 0.5 to increase positive samples in minibatches.
Results
On INbreast, the fully automated system achieved AUC = 0.95 for breast-level classification. Detection evaluation used lesion-level FROC analysis, showing 90% sensitivity at 0.3 false positive marks per image and complete lesion detection at 3 false positive marks per image.
- Classification performance: AUC = 0.95 was achieved on the public INbreast dataset for breast-level classification.The 95 percentile interval was 0.91 to 0.98, estimated from 10000 bootstrap samples.
- Detection performance: Detection performance was evaluated with a Free-response ROC curve, measuring lesion-level sensitivity against false positive marks per image.A detection counted as correct when the proposed lesion center fell inside a ground truth box; the DM challenge dataset lacked lesion annotations for FROC analysis.
- Detection errors: At sensitivity = 0.9 and 0.3 false positive marks per image, the detector’s examples included correctly classified, false positive, and missed malignant lesions.Most false positive detections were benign masses or calcifications.
- Detection errors: 3 false positive marks per image yielded correct detection of all lesions, whereas 10% of ground truth malignant lesions were missed at the lower detection threshold.The missed lesions were not completely overlooked by the model.
Discussion
The Faster R-CNN CAD system achieved strong mammogram lesion classification and detection performance without annotations or user interaction. Its detection evaluation was limited by the small size of the pixel-level annotated INbreast dataset.
- AUC = 0.85 earned the method 2nd position in the Digital Mammography DREAM Challenge on the final validation dataset.
- 90% malignant-lesion sensitivity was achieved on INbreast with 0.3 false positive marks per image.The system used mammograms as its only input, without annotation or user interaction.
- The results indicate that deep learning object detection can perform lesion detection on mammograms and may replace expensive traditional CAD methods.The proposed approach could potentially help radiologists detect more cancers as a perception enhancer tool.
- The detection performance could only be evaluated on the small INbreast dataset because few publicly available datasets contain pixel-level annotations.Classification performance was evaluated on a large screening dataset, but detection evaluation was limited to INbreast.