Author:

Keywords:

PSI_MIC

Abstract:

Brain tumor lesions, MS white matter lesions and ischemic stroke lesions are all similar in the sense that they appear in MR images as a collection of spatially coherent voxels that are hyper-intense, hypo-intense or a combination of these with respect to the healthy tissue. Moreover, these lesions evolve relatively smoothly over time. This work presents a collection of methods that tackle the problems of single-temporal segmentation, multi-temporal segmentation and multi-temporal regression of different types of lesions in brain MR images. This collection of methods will be referred to as TIMinG, a toolkit for Tumor Image-based Morphology and Growth. As the name implies, TIMinG is originally designed for brain tumors. With only some minor modifications, we show that TIMinG is also able to perform single-temporal segmentation and multi-temporal segmentation of MS lesions as well as single-temporal segmentation of stroke lesions. Basically, TIMinG is a novel fully-automated lesion segmentation method that is directly applicable to any individual patient MR image. The method is independent of the scanner or acquisition protocol and does not require any manually annotated training data. TIMinG is the best performing untrained whole tumor segmentation method and the third best overall method (trained and untrained), compared with other methods from the MICCAI Brain Tumor Segmentation 2012-2013 Challenge. Moreover, TIMinG is able to segment MS lesions and stroke lesions with a performance that is similar to state-of-the-art dedicated MS lesion and stroke lesion segmentation methods. The fact that the method is untrained and unsupervised and generalizes well to different MR modalities and various types of lesions makes the method well suited for lesion quantification in small clinical settings and pre-clinical research settings. Secondly, TIMinG is able to segment all images in a patient time series simultaneously. The TIMinG multi-temporal segmentation method is a mathematical generalization of the single-temporal method. The method is validated for both tumor and MS lesion time series. The overall segmentation accuracy and consistency is compared with the single-temporal segmentation results. It is shown for tumor patient time series that the segmentation accuracy is increased. Moreover, an increase in the temporal consistency of the tumor segmentations is shown, i.e. the TIMinG multi-temporal segmentation method performs better than the single-temporal method in measuring the whole tumor volume changes between consecutive time points. Finally, TIMinG models the lesion growth over time and simulates its shape at every moment. Simultaneously, it estimates the velocity of the local lesion growth. The real clinical acquisition dates are hereby taken into account. TIMinG's multi-temporal regression is illustrated on 16 brain tumor patient time series and the model prediction power is validated using a leave-one-out cross-validation approach. Overall, TIMinG is intended to be used by clinicians as an exploratory tool to assess the lesion morphology and growth. It can also serve as a basis for subsequent quantification, like maximal growth velocity, principal axis of growth, local acceleration of the lesion, etc. It suffices for the clinician to have a time series -two or more time points- of MR images available to obtain quantitative measures about the lesion morphology and growth.