Detecting and classifying lesions in mammograms with Deep Learning.

Ribli D, Horváth A, Unger Z, Pollner P, Csabai I. · 2018

Read Detecting and classifying lesions in mammograms with Deep Learning. by Ribli D, Horváth A, Unger Z, Pollner P, Csabai I. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5854668.txt.

Introduction

Ribli and co-authors present a deep learning system for mammogram reading. Their article appeared in Scientific Reports. The system finds lesions in breast X-rays and labels them benign or malignant. It needs no human input once an image is loaded. The team built it on the Faster R-CNN object detection framework. Older computer aids gave mixed gains for radiologists, they note.

The authors report top results on the public INbreast database. Their system scored an AUC of 0.95 there. It also took second place in the Digital Mammography DREAM Challenge. As a detector, it missed few cancers and raised few false marks. Code, the trained model, and a viewer plugin are shared online.

Catalog record

CREATORRibli D, Horváth A, Unger Z, Pollner P, Csabai I.
YEAR2018
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5854668

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Detecting and classifying lesions in mammograms with Deep Learning By Dezső Ribli, Anna Horváth, Zsuzsa Unger, Péter Pollner, István Csabai Published in: Scientific Reports DOI: 10.1038/s41598-018-22437-z ABSTRACT Abstract In the last two decades, Computer Aided Detection (CAD) systems were developed to help radiologists analyse screening mammograms, however benefits of current CAD technologies appear to be contradictory, therefore 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 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 published online at https://github.com/riblidezso/frcnn_cad. INTRODUCTION Screening mammography Breast cancer is the most common cancer in women and it is the main cause of death from cancer among women in the world1. Screening mammography has been shown to reduce breast cancer mortality by 38–48% among participants2. In the EU 25 of the 28 member states are planning, piloting or implementing screening programs to diagnose and treat breast cancer in an early stage3. During a standard mammographic screening examination, X-ray images are captured from 2 angles of each breast. These images are inspected for malignant lesions by one or two experienced radiologists. Suspicious cases are called back for further diagnostic evaluation. Screening mammograms are evaluated by human readers. The reading process is monotonous, tiring, lengthy, costly and most importantly, prone to errors. Multiple studies have shown that 20–30% of diagnosed cancers could be found retrospectively on the previous negative screening exam by blinded reviewers4–9. The problem of missed cancers still persists despite modern full field digital mammography (FFDM)4,8. The sensitivity and specificity of screening mammography is reported to be between 77–87% and 89–97% respectively. These metrics describe the average performance of readers, and there is substantial variance in the performance of individual physicians, with reported false positive rates between 1–29%, and sensitivities between 29–97%10–12. Double reading was found to improve the performance of mammographic evaluation and it has been implemented in many countries13. Multiple readings can further improve diagnostic performance up to more than 10 readers, proving that there is room for improvement in mammogram evaluation beyond double reading14. Computer-aided detection in mammographic screening Computer-aided detection (CAD) solutions were developed to help radiologists in reading mammograms. These programs usually analyse a mammogram and mark the suspicious regions, which should be reviewed by the radiologist15. The technology was approved by the FDA and had spread quickly. By 2008, in the US, 74% of all screening mammograms in the Medicare population were interpreted with CAD, however the cost of CAD usage is over $400 million a year11. The benefits of using CAD are controversial. Initially several studies have shown promising results with CAD6,16–20. A large clinical trial in the United Kingdom has shown that single reading with CAD assistance has similar performance to double reading21. However, in the last decade multiple studies concluded that currently used CAD technologies do not improve the performance of radiologists in everyday practice in the United States11,22,23. These controversial results indicate that CAD systems need to be improved before radiologists can ultimately benefit from using the technology in everyday practice. Currently used CAD approaches are based on describing an X-ray image with meticulously designed hand crafted features, and machine learning for classification on top of these features15,24–27. In the field of computer vision, since 2012, deep convolutional neural networks (CNN) have significantly outperformed these traditional methods28. Deep CNN-s have reached, or even surpassed, human performance in image classification and object detection29. These models have tremendous potential in medical image analysis. Several studies have attempted to apply Deep Learning to analyse mammograms27,30–32, but the problem is still far from being solved. The Digital Mammography DREAM Challenge The Digital Mammography DREAM Challenge (DM challenge)33,34 asked participants to write algorithms which can predict whether a breast in a screening mammography exam will be diagnosed with cancer. The dataset consisted of 86000 exams, with no pixel level annotation, only a binary label indicating whether breast cancer was diagnosed within the next 12 months after the exam. Each side of the breasts were treated as separate cases that we will call breast-level prediction in this paper. The participants had to upload their programs to a secure cloud platform, and were not able to download or view the images, nor interact with their program during training or testing. The DM challenge provided an excellent opportunity to compare the performance of competing methods in a controlled and fair way instead of self-reported evaluations on different or proprietary datasets. MATERIAL AND METHODS Data Mammograms with pixel level annotations were needed to train a lesion detector and test the classification and localisation performance. We trained the model on the public Digital Database for Screening Mammography (DDSM)35 and a dataset from the Semmelweis University in Budapest, and tested it on the public INbreast36 dataset. The images used for training contain either histologically proven cancers or benign lesions which were recalled for further examinations, but later turned out to be nonmalignant. We expected that training with both kinds of lesions will help our model to find more lesions of interest, and differentiate between malignant and benign examples. The DDSM dataset contains 2620 digitised film-screen screening mammography exams, with pixel-level ground truth annotation of lesions. Cancerous lesions have histological proof. We have only used the DDSM database for training our model and not evaluating it. The quality of digitised film-screen mammograms is not as good as full field digital mammograms therefore evaluation on these cases is not relevant. We have converted the lossless jpeg images to png format, mapped the pixel values to optical density using calibration functions from the DDSM website, and rescaled the pixel values to the 0–255 range. The dataset from the Department of Radiology at Semmelweis University in Budapest, Hungary contains 847 FFDM images of 214 exams from 174 patients, recorded with a Hologic LORAD Selenia device. Institutional board approval was obtained for the dataset. This dataset was not available for the full period of the DM challenge, therefore it was only used for improvement in the second stage of the DM challenge, after pixel level annotation by the authors. The INbreast dataset contains 115 FFDM cases with pixel-level ground truth annotations, and histological proof for cancers36. We adapted the INbreast pixel level annotations to suit our testing scenario. We ignored all benign annotations, and converted the malignant lesion annotations to bounding boxes. We excluded 8 exams which had unspecified other findings, artefacts, previous surgeries, or ambiguous pathological outcome. The images had low contrast, therefore

Read the complete text