Skip to main content
European Commission logo print header

Improved detection and characterisation of breast cancer using multimodal<br/>magnetic resonance imaging and novel computer-aided<br/>detection/evaluation (CADe) techniques

Final Report Summary - CADE4BMRI (Improved detection and characterisation of breast cancer using multimodal<br/>magnetic resonance imaging and novel computer-aided<br/>detection/evaluation (CADe) techniques)


Magnetic resonance imaging (MRI) is being increasingly used clinically as a supplemental tool for the detection and characterisation of breast cancer. A major limitation of breast MRI is that although it has a high sensitivity to breast cancer, its specificity is poor to moderate (i.e. it detects too many false positives). This not only leads to patient anxiety and unnecessary interventions but also is costly.

The aim of this research was to improve the specificity of breast MRI for the detection of breast cancer, and therefore the clinical utility of the technique, by:

1. Integrating information from multiple MRI techniques—T1- and T2-weighted anatomical images, dynamic contrast-enhanced (DCE) MRI, and diffusion weighted imaging (DWI)—to better characterise the properties (blood perfusion, morphology, tissue microstructure) of breast lesions; and
2. Reducing the subjectivity in routine clinical interpretation of the MRI data through the use of novel computer visualisation and quantitative image analysis techniques.

To this end the research had the following objectives/tasks (abridged):

Task 1. To obtain multi-technique MRI data from 100 routine clinical breast MRI examinations (in which the reporting radiologist identified at least one suspicious lesion whose status was confirmed by biopsy and microscopic examination), and delineate and label each lesion.
Task 2. To evaluate/develop computer-based methods to correct for patient movement during acquisition of the MRI data, spatially align the data, and perform bias field correction (the bias field manifests as a corrupting low frequency signal variation in the MR image data).
Task 3. To evaluate/develop several mathematical models for quantitatively characterising the pattern of wash-in and wash-out of contrast agent observed in the DCE-MRI data (3D images acquired before and several times after the injection of a gadolinium-containing contrast agent). This is also useful for visualising and automatically segmenting (i.e. locating and delineating) suspicious lesions.
Task 4. To develop a 3D colour-coded visualisation of the 4D DCE-MRI data that the radiologist can interact with using a stereo display and haptic device (permitting the radiologist to physically probe and “feel” tissue properties via tactile feedback and to interactively segment suspicious tissue).
Task 5. To have a radiologist at Sahlgrenska University hospital evaluate the complete system developed in Task 4.
Task 6. To implement both conventional and novel quantitative features (measurements) for characterising lesions.
Task 7. To compute the features from Task 6 for all of the data from Task 1 and to determine the best subset of features for classifying lesions as either benign or malignant.
Task 8. To estimate the expected performance of the classifier of Task 7 on unseen data.


Note: The project was terminated 11 months early because the fellow accepted an appointment at the University of Western Australia. Consequently all tasks were completed except tasks 3 and 4.

1. We developed a method for automatically selecting a region of apparent diffusion coefficient (ADC) hypointensity in suspicious lesions. The ADC is computed from the DWI data and is a measure of the diffusion of water molecules in tissues. It is affected by changes in the microstructure and local tissue architecture of the tissues due to disease. Our method was presented at the 20th Scientific Meeting and Exhibition of the International Society for Magnetic Resonance in Medicine held from 5-11 May 2012 in Melbourne Australia.
2. We developed a method for thresholding (isolating tissue of interest from background) two or more images acquired with different imaging modalities (e.g. PET, CT, MRI) or acquisition protocols (e.g. DCE-MRI, DWI). This was published in the journal, IET Computer Vision in June 2013.
3. We developed the first algorithm for fully automatic detection and delineation of tissue suspicious for malignancy in breast MRI data. This was published in the Journal of Magnetic Resonance Imaging in April 2014.
4. We developed/implemented a set of new quantitative features (measurements) based on qualitative descriptors from the American College of Radiology’s BI-RADS® (Breast Imaging-Reporting and Data System) Atlas and Kaiser’s “Signs in MR-Mammography” (Springer Berlin Heidelberg, 2008). We have submitted a paper on this work to the Journal of Medical Imaging and have been requested to do a major revision.
5. We have developed new supervoxel-based algorithms for segmenting the breast tissue in MR images and for extracting features (quantitative measurements) from suspicious lesions. This work is detailed in Chalmers MSc Thesis EX045/2014.

Fellow’s career development/integration
1. The fellow was awarded the academic title of Docent (Associate Professor) on 2014-06-03 after his application passed external expert review and his Docent lecture, entitled “Improved detection and characterisation of breast cancer using multi-modal magnetic resonance imaging and novel computer-aided detection/diagnosis (CAD) techniques”, was approved by the Head of the Department of Signals and Systems at Chalmers University of Technology. This qualification required, in addition to a PhD, substantially greater documented independent ability to lead in formulating and solving scientific research problems, as well as pedagogic competence at the advanced graduate level of study.
2. The fellow has broadened his teaching experience by teaching into two graduate courses (Image Analysis and Diagnostic Imaging) in the Department of Signals and Systems at Chalmers in both 2013 and 2014. He additionally had the role of coordinator for the second course.
3. The fellow successfully supervised an MSc thesis, entitled “Supervoxel-based algorithms for use in breast MRI CAD systems”, based on this project.
4. The fellow successfully supervised (as a joint principal supervisor) a PhD thesis in the School of ITEE at the University of Queensland, Australia, entitled “Computer assisted detection and characterisation of breast cancer in MRI”. This was possible because he holds an adjunct (unpaid) appointment in the School of ITEE. The thesis passed examination, subject to minor corrections, on 24/10/2014.
5. The fellow applied for an appointment as Associate Professor at the University of Western Australia in July 2014. He was offered the position in October and accepted. His last working day at the host institution (Chalmers University of Technology) was 24 October 2014. Thus this project had to be terminated early (31 of 42 months).


The principal outcomes from this project are several algorithms/methods for identifying, quantitatively characterising, and classifying (as benign or malignant) suspicious lesions in image data from a breast MRI examination. These have been published in refereed journal and conference papers, a PhD thesis, and an MSc thesis and would be directly relevant to developers of MRI CAD systems.

The final results (PhD thesis) demonstrate that the feature set (set of quantitative measurements) and the computer classifier(s) developed during this study yield improved sensitivity and specificity for the detection of breast cancer. This offers the possibility for sensitive and specific detection of early breast cancer and concomitant reduction in morbidity and mortality from breast cancer. This is not only of interest to the EU but globally.