Redona Brahimetaj is a PhD researcher at the department of Electronics and Informatics (ETRO) at the Vrije Universiteit Brussel. She is interested in developing Computer Aided Detection and Diagnosis system to early detect breast cancer based only on high resolution 3D microcalcifcation images.
Breast cancer is one of the most common cancers in women worldwide. As the tumor itself is often not clearly visible in mammograms, the main indicator of an early non-palpable breast cancer is the presence of microcalcifications, small calcium deposits which are often found in breast tissue. Although microcalcifications may be an early indicator of breast cancer, they are not always restricted to malignancies as they are also present in common benign lesions.
Early detection of MCs and analysis of their properties, are important for the overall prognosis and management of breast cancer. Discriminating between MCs found in benign and malignant lesions based on their appearance on mammograms is in many cases difficult for radiologists considering the limitations that the current standard screening modalities used (such us mammography) have nowadays (i.e.: lack of contrast, low resolution, etc). Although mammography provides one of the highest spatial resolution among the imaging modalities used to detect breast cancer, its resolution still remains low to discriminate between benign and malignant individual MCs. Moreover, as mammography is a 2D projection, the superposition of tissue can hide MCs and alter their appearance depending on their orientation relative to the image plane.
In order to assist radiologists in detecting and characterising tumor types in different image modalities, Computer Aided Detection (CAD) systems have been developed. Despite the promising results achieved over the years, the current existing CAD systems (a) suffer from a high number of false positive and false negative rates and, (b) mostly study MCs properties in 2D or 3D low resolution images. Since the most realistic way to characterise a 3D object is to use a high-resolution 3D imaging modality, we focus on studying MCs which are visualised in 3D using a high resolution scanner. As in vivo imaging is still not possible, the images used are obtained from biopsies and scanned using a micro-CT scanner. Given the evidence on the link between image features of MCs and malignancy and tumor type, we already could conceive a CAD systems with both a low false positive and false negative rate.
Given the important legacy of anatomical and pathological knowledge, we will evaluate whether deep learning methods enriched with hand-crafted features can outperform a purely deep learning method. We hypothesise that such a system will be able to discriminate between healthy patients and patients suffering from breast cancer, but also will be able to classify malignant MCs into different breast tumor types, with less false positives and false negatives than the current systems not taking into count individual MCs. Although our research can not currently be applicable in vivo (3D high resolution in vivo breast imaging is still not possible), its potential for use in intra operative imaging to reduce the waiting time between tissue extraction and anatomopathological results, is high. As a long term achievement, with our study we aim to avoid unnecessary biopsies and considerably reduce costs for the healthcare system.