- This event has passed.
BioMedIA Talk: Thijs Kooi (Radboud University Medical Center)
February 23 @ 1:00 pm - 2:00 pm
Computer aided diagnosis of breast cancer in mammography using deep neural network architectures
In spite of significant advances in the past decades, breast cancer remains a lethal disease, with an estimated 40000 death in the US for 2016. Early detection due to screening has played a key role in reducing mortality, but tumors are still missed and women are referred unnecessarily for treatment of benign abnormalities. To improve screening, Computer Aided Diagnosis (CAD) systems are being developed to aid and ultimately replace human readers.
In a recent paper, we showed that a CAD system based on a deep Convolutional Neural Network (CNN) can operate at the level of expert radiologists when classifying Regions of Interest (ROI). However, a mammographic case is a complex structure conveying temporal, multi-view and symmetry aspects. Additionally, benign mass-like abnormalities such as cysts are often confused with malignant lesions, both by radiologists and the CNN. To outperform human readers on a case level, all these sources of information need to be parsed and false positives need to be reduced. This talk will cover several deep learning approaches to tackle this and work on a Conditional Random Field (CRF) trained on top of the CNNs to model interactions between ROIs and integrate all sources of information.