Xuanhe Zhao Robert N Noyce Career, Development Associate Professor of Mechanical Engineering
Alberto Rodriguez Walter Henry Gale (1929) Career Development Assistant Professor of Mechanical Engineering
Miro Kazakoff Lecturer, Work and Organization Studies, MIT Sloan School of Management
David Simchi-Levi
Professor of Civil and Environmental Engineering and Engineering Systems; Head, Accenture and MIT Alliance in Business Analytics
Donna Rhodes
Principal Research Scientist in the Sociotechnical Research Center; Director of the Systems Engineering Advancement Research Initiative (SEAri)
Sanjay Sarma Vice President of Open Learning; Professor of Mechanical Engineering
Currently, medical images require a physician to extract clinically relevant information. This talk will explore current work towards making images part of the quantitative medical history and to enable large-scale image-based studies of disease. Although large databases of clinical images contain a wealth of information, medical acquisition constraints result in sparse scans that miss much of the anatomy. These characteristics often render computational analysis impractical as standard processing algorithms tend to fail when applied to such images. Our goal is to enable application of existing algorithms that were originally developed for high resolution research scans to severely undersampled images. Application of the method is illustrated in the context of neurodegeneration and white matter disease studies in stroke patients.
This talk introduces a new generation of machine learning methods that provide state of the art performance and are very interpretable, introducing optimal classification (OCT) and regression (ORT) trees for prediction and prescription with and without hyperplanes. This talk shows that (a) Trees are very interpretable, (b) They can be calculated in large scale in practical times, and (c) In a large collection of real world data sets, they give comparable or better performance than random forests or boosted trees. Their prescriptive counterparts have a significant edge on interpretability and comparable or better performance than causal forests. Finally, we show that optimal trees with hyperplanes have at least as much modeling power as (feedforward, convolutional, and recurrent) neural networks and comparable performance in a variety of real world data sets. These results suggest that optimal trees are interpretable, practical to compute in large scale, and provide state of the art performance compared to black box methods.