Active learning based intervertebral disk classification combining shape and texture similarities

We consider the problem of the computer aided diagnosis on intervertebral disk degeneration in this paper, where a framework of classifying the disks into healthy and degenerated groups is proposed. First, we propose to separately model the shape and texture similarities and combine them into a single measure to describe the difference of imaging appearance between healthy and degenerated disks. Then we introduce the active learning strategy into our framework, aiming at training a well performed classifier with only a subset of data that are most representative. This would save much human effort by avoiding labeling all training data manually. Herein we propose a simple but effective improved transductive experimental design, taking advantage of the latent structure of the data, to achieve the automatic data selection. Experiments on synthetic data and real disc MRI show the effectiveness of our approach.

► We introduce the active learning strategy into the CAD on the inter-vertebral disk degeneration.
► The imaging appearance is characterized by combining the shape and texture strength.
► We propose an improved transductive experimental design to select representative data.