Meeting Banner
Abstract #4030

Do we still need mathematical modeling in the age of deep learning? A case-study comparison of the Tofts model versus end-to-end deep learning in prostate cancer segmentation

Alessandro Guida1, Peter Q Lee2, Steve Patterson3, Thomas Trappenberg2, Chris V Bowen1,4, Steven Beyea1,4, Jennifer Merrimen5, Cheng Wang5, and Sharon E Clarke4

1Biomedical Translational Imaging Centre, Halifax, NS, Canada, 2Faculty of Computer Science, Dalhousie University, Halifax, NS, Canada, 3Nova Scotia Health Research Foundation, Halifax, NS, Canada, 4Department of Diagnostic Radiology, Dalhousie University, Halifax, NS, Canada, 5Department of Pathology, Dalhousie University, Halifax, NS, Canada

The rise in popularity of deep learning is revolutionizing the way biomedical images are acquired, processed, analyzed. Just a few years ago, extracting high-level understanding from biomedical images was a process restricted to highly trained professionals often requiring multidisciplinary collaborations. In the work presented, we showcase a study that compares the performance of a model trained end-to-end using a novel deep learning architecture, versus a model trained on its corresponding state-of-the-art mathematically engineered feature. Results show that end-to-end deep learning significantly outperforms the mathematical model, suggesting that feature engineering will play a less important role in the coming years.

This abstract and the presentation materials are available to members only; a login is required.

Join Here