School of Computing Research Colloquia

Visual Parameter Optimization for Biomedical Image Analysis: A Case Study

Hannes Pretorius , Visualization and Virtual Reality research group 

The conventional approach for parameter optimization of biomedical image analysis algorithms is to tweak parameters by trial-and-error. This presents a challenge: parameter space is often inadequately explored and, consequently, output quality suffers. Interactive visualization can alleviate this problem but has not been widely adopted. Moreover, few examples of the successful application of visualization for parameter optimization of image analysis algorithms have been published. To address this and to illustrate the potential usefulness of interactive visualization, we present a case study.

A multidisciplinary team developing novel image segmentation software for histopathology was observed. Within the context of our study, our hypotheses were confirmed: (1) using interactive visualisation, participants considered larger parts of parameter space than they had previously by trial-and-error; (2) participants gained a better understanding of their algorithm (an unknown logic error and errors in its implementation were discovered); and (3) participants achieved higher quality output. Our work is also an example of the value of case studies in iterative design. We describe how a valuable additional requirement was revealed (the importance of derived measures) and how our visualization method was extended to cater for this.