A feasibility study on automated protein aggregate characterization utilizing a hybrid classification model

Abstract

The characterization of cytoplasmic protein aggregates based on time-lapse fluorescence microscopy imaging data is important for research in neuro-degenerative diseases such as Parkinson. As the manual assessment is time-consuming and subject to significant variability, incentive for the development of an objective automated system is provided. We propose and evaluate a pipeline consisting of cell-segmentation, tracking and classification of neurological cells. Focus is specifically on the novel and challenging classification task which is covered by relying on feature extraction followed by a hybrid classification approach incorporating a support vector machine focusing on mainly stationary information and a hidden Markov model to incorporate temporal context. Several image representations are experimentally evaluated to identify cell properties that are important for discrimination. Relying on the proposed approach, classification accuracies up to 80 % are reached. By extensively analyzing the outcomes, we discuss about strengths and weaknesses of our method as a quantitative assessment tool.