Hajim School of Engineering and Applied Sciences UofR logo BME SMD logo Hajim SEAS logo

Thursday, Aug 08, 2013

2:00 PM3:00 PM MC 2-6408 (K-207)

Upcoming Events

There are no upcoming events to display

Past events

BME PhD Thesis Defense Seminar: Mahesh Nagarajan

A Framework for Computer-Aided Diagnosis with Novel Computational Methods for Characterizing Healthy and Pathological Soft Tissue Patterns on Medical Images

Professor Axel Wismueller, M.D., Ph.D.


Computer-aided diagnosis (CADx) can assist radiologists in extracting and interpreting clinically relevant findings from suspicious patterns on medical images in a consistent manner. CADx is currently implemented as image pattern processing pipeline with three components - feature extraction (pattern characterization), feature reduction (efficient representation of feature set) and pattern classification. This work presents an enhanced CADx methodology with improvements to both feature extraction and feature reduction components. Characterization of healthy and pathological patterns on medical images through topological features derived from Minkowski Functionals and geometrical approaches derived from the Scaling Index Method (SIM) is explored as an alternative to conventionally used statistical features derived from gray-level co-occurrence matrices. CADx integration of dimension reduction in conjunction with out-of-sample extension is investigated to address the shortcoming of previous attempts where such integration violated training-test set separation requirements of the supervised learning step. The improved CADx methodology presented in this work was demonstrated using two medical datasets - (1) the classification of small diagnostically challenging lesions on dynamic contrast-enhanced breast MRI and (2) the classification of healthy and osteoarthritic cartilage on phase contrast CT images of the knee patella. The results lend credence to the ability of CADx, as implemented in this work, to characterize soft tissue patterns on medical images in an automated and non-subjective manner for purposes of classification between healthy and diseased patterns. Such work can be of significant interest for identifying imaging markers that can assist radiologists in extracting clinically relevant findings, and perhaps even tracking disease progression and response to therapeutic intervention.