דלג לתוכן (מקש קיצור 's')
אירועים

אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

event speaker icon
רמי בן-ארי (IBMׁ)
event date icon
יום שלישי, 09.01.2018, 11:30
event location icon
חדר 337, בניין טאוב למדעי המחשב
In several clinical routines such as mammography, the early detection of breast cancer has an enormous impact in the patient survival. Radiologists nowadays are overwhelmed by the imaging data coming from numerous screening and diagnostic tools. This puts pressure on the early detection of breast cancer has an enormous impact in the patient survival. Radiologists nowadays are overwhelmed by the imaging data coming from numerous screening and diagnostic tools. This puts pressure on the radiologists to be highly sensitive while keeping false positive rate low. Cognitive systems in computer aided diagnosis strive for becoming a second reader on detecting malignancies in breast imaging. However, standard detection approaches require a hard labor of annotation of the findings in the imagery data sets in order to train a “fully” supervised model. This is a tedious task, which is often out of reach due to the high cost and unavailability of expert clinicians. In this talk I’ll present two recently published methods for supervised learning, where the mammogram (MG) images are tagged only on at the global level, also known as “weakly labeled”, i.e. without local annotations. The weak tags on MG data can be extracted from the retrospective medical records without the need for radiologist review and annotation. In the first approach I’ll describe our deep learning method for classification of whole mammograms according to the type of lesions contained in these large images. In addition to the classification method, we suggest a novel loss function that directly optimizes the Area under the ROC curve, as commonly the positive and negative classes are highly imbalanced. Our second method, further tackles the localization of the malignant findings as the local source of discrimination between classes, without using any local annotations in training. As the demand for Big Data continuous to grow causing the annotation labor by expert clinicians a major obstacle for traditional fully supervised learning, the weakly labeled approach thus promises a rescue from this tangle.