Skip to main content

Intelligent Automated System for detecting Diagnostically Challenging Breast Cancers

Periodic Reporting for period 1 - SmartMammaCAD (Intelligent Automated System for detecting Diagnostically Challenging Breast Cancers)

Reporting period: 2015-09-01 to 2017-08-31

"To run a chance of surviving breast cancer it is uttermost important to discover malignant tumours at an early stage. Deaths by breast cancer are highly reduced by early treatment. In his fundamental publication “Signs In MR-Mammography” Werner A. Kaiser states: “If we had a diagnostic method that enabled us to detect and remove all breast cancers 5 to 10 mm in size, we could practically eliminate breast cancer deaths”. Methods able to diagnose even very small lesions play an important role in the fight against breast cancer.

Now days, computers have dramatically increased their computation abilities. With the rise of artificial intelligence, high demanding tasks can be achieved in a matter of seconds with the aid of new machine learning algorithms. Also, the field of medical imaging has notably contributed to assist physicians with novel and non-invasive techniques for diagnosis and prognosis, through high resolution images of the internal organs by magnetic resonance imaging (MRI), or nuclear imaging. The main goal and overall objective of this project is to develop computer aided diagnosis (CAD) methods, and image processing techniques to improve diagnostic accuracy and efficiency of cancer-related breast lesions. CAD systems will aid physicians by detecting challenging lesions with high accuracy, ""learning"" to detect them from previously diagnosed cases. Those challenging cases are usually non-mass enhancing lesions, not easy to detect by direct methods, or requiring invasive methods to diagnose them, such as biopsies. Therefore, the use of CAD systems in the clinical practice is not only important for reducing the impact of the most aggressive breast cancers, but also to increase the patient care and comfort."
In this first phase of the project, in collaboration with the Florida State University, two main results have been achieved:

1) The complete processing of a dataset of several hundreds of DCE-MRI images, each of them containing more than a milllion voxels, including motion compensation and segmentation.

This first objective of the project has been achieved with the active participation of the Florida State University, which provided a dataset of Dynamic Contrast Enhanced (DCE) Magnetic Resonance Images (MRI) of patients diagnosed with non-mass enhanced lesions (NMHL). The dynamic behaviour of NMHL in DCE-MRI overlaps with non malign tissue, and constitutes a challenge for diagnosis. Therefore, accurate preprocessing methods are required for motion compensation and segmentation, decreasing the impact that noise, enhancing of internal organs, movements and misalignments produce on relevant signals. I performed a SPM based registration on the whole dataset, studying rigid and non-rigid registration. The result improved the original affine registration that had been only tested before on the dataset. I anonymized the database and uniformized the different protocols available in the database, concretely a spatial high resolution 3T, temporal high resolution dataset with a 1.5T 5 temporal point dataset.

For segmentation, I developed a new algorithm based on 3D Gabor filtering that is able to automatically detect wall-like structures in medical imaging, therefore isolating breast dynamic signals from noisy sources separated by the chest wall. The implementation of the segmentation algorithm was made available through the collaborative-code platform github, under the GPL-3.0+ license, and constitutes one of the building blocks of the SmarMammaCAD software. It will be presented in future conferences and journals.

2) Preliminary studies, theoretical and practical, performed on this database to identify salient features for malign lesion identification and classification using automatic methods as independent component analysis (IAC) or support vector machines (SVM).

Regarding theoretical studies, the problem of semisupervised learning was studied from a case-based learning paradigm perspective and a hypothesis testing, respectively. When statistical testing hypothesis are introduced in the feature selection scheme from a pool of non-labeled data, an improvement on performance parameters is obtained, such as cross-validated estimated accuracy and error rates, presented at the SPIE 2017 conference and published in:
JM Gorriz, et. al Expert Systems with Applications 2017 90, 40-49
JM Gorriz, et. al IEEE Access 2017/5

In practical applications, and under the hypothesis of expressiveness of the dynamic behaviour of malign tissues, I employed blind sources separation techniques as ICA to produce an enriched characterization of the signals that improved the separability between tissues, in contrast with other dynamic characterizations as the 3-point method, as well as machine learning techniques to automatically separate and classify dynamic curves at a voxel level. I tested this approach in a subset of the processed dataset with promising results presented in the ECR 2017:

I.A. Illán et. al. Machine learning for challenging tumour detection and classification in breast cancer. ECR 2017 / C-3151

Correctly validated, this software will constitute the second building block of the SmartMammaCAD software.
For the correct segmentation and isolation of breast tissues from other organs imaged, a novel techinque has been developed that is specially suitable for identifying wall-like structures in medical imaging. This technique has the potential to be used in many other important problems in medical imaging, such as the identification of the hemisphere dividing plane in the brain, which is very relevant for studies on the left-right hemisphere symmetry and it's effects (handedness, gender, language,...) and other problems requiring the identification of wall-like organic structures.

Moreover, blind source separation techinques have been proved to be usefull in the analysis and identification of relevant kinetic features for malignant tissues classification. This prove-of-concept study will be trained and validated on a big dataset of clinical practice cases, with an open source implementation for clinical use, a pilot tool developed in collaboration with radiologists. This tool will be optimized for detecting non-mass enhancing lessions that usually require biopsies for diagnosis, reducing drastically the impact on the patients treatment when the desired accuracy is achieved and the effectiveness is granted.
Example of malign tissue identification by an automated method