Motivation: Mammography is one of the most widely used techniques today to screen for breast cancer. Due to mammographic screening’s manual nature many breast masses are misdiagnosed or missed completely. And achieving high diagnostic accuracy requires expertise acquired over many years of experience as a radiologist.
Current Study: A considerable amount of research has been done on computer aided detection’s (CAD) effects on better helping detect early breast cancer. However, Many drawbacks still exist which are based off:
(1). Irregular breast tissue lesions can be very small, and the usual classification frameworks with small image sizes reduces the resolution of the image making it near to impossible to be able to see them
(2). Pixel level annotations contain far more information than simple exam level annotations, especially for small objects (such as lesions)
(3). The clinical usability of a lesion detector is much more practical than a simple classifier, which does not tell a radiologist where to look or focus on.
Objective: More importantly, current researches in this area aim to use CAD to automatically distinguish malignant and benign only based on Mammogram, replace radiologists and ignore the others diagnose methods (e.g., biopsy). However, the current AI technology still cannot compete with human. Therefore, here we present an approach to better help detect early breast cancer by developing a software that acts as a ‘second-opinion’ to the radiologist without compromising their medical expertise.
Figure 1: System Diagram. A deep learning model is trained to recognize a small region of the whole image, i.e. image patches. Irregular breast lesions are extracted through these patch images. Then the trained deep learning model is used to "scan" a whole image and make predictions for all image patches and better highlight areas of concern and irregularity.
Figure 2: ROI Detector. A deep learning model (YOLO v3) is trained to recognize a small irregular region of the whole image, i.e. image patches. Uses logistic regression and a threshold to predict multiple labels for one image. There are 21 classes (irregular shapes) in total (13 calcification and 8 mass)
Comments