Computational Methods for Biomedical Data

Abstract

Biomedical data provides a rich opportunity for the development of new statistical and deep learning tools. In this work we focus on two problems in biomedical data analysis, namely classification of effusion synovitis in knee MRI data and data adaptive dose finding studies in clinical trials. The classification of effusion synovitis (ES) in knee MRI data is a challenging problem due to the inherent nature of the data, where only a small percentage of the image is indicative of ES. As a result, the usage of popular image analysis algorithms produces poor classification performance in both the single and multi-slice case. Our work introduces a graph convolutional network (GCN) model to combine information both within and between-slice for MRI series to accurately classify ES. We first develop what we term the ‘single-slice’ model, where image networks are constructed from only one slice per MRI series and are used as input to our GCN. Our method outperforms commonly used existing models, with an area under the curve of 80.95 percent, and a Matthews correlation coefficient of 60.8. Next, we extend our method by constructing image networks across slices within an MRI series, which we term the ‘multi-slice’ model, allowing our method to reflect clinical practice. This approach increases classification performance and outperforms existing methods by a larger margin, with an area under the curve of 86.11 and a Matthews correlation coefficient of 57.5. Lastly, we turn our attention to dose finding in a phase 2 group sequential trial. We introduce the usage of a group-specific semi-parametric model that is formulated to select the model which minimizes prediction error given all observed data and assign doses to incoming groups. We show that our method results in a lower overall error for different types of phase 2 data when compared to the fully parametric and fully non-parametric model alone.