3D Convolutional Neural Network for Chronic Stroke Lesion Segmentation
|Title||3D Convolutional Neural Network for Chronic Stroke Lesion Segmentation|
|Publication Type||Conference Paper|
|Year of Publication||2016|
|Authors||Wang, Y., X. Wang, A. K. Katsaggelos, and T. B. Parrish|
|Conference Name||The Organization for Human Brain Mapping (OHBM)|
|Conference Location||Geneva, Switzerland|
Stroke is one of the leading causes of death and disability. It is important to have accurate diagnosis for proper treatment as well as good prognosis to help the patients prepare for rehabilitation. Brain MRI scan is used for the identification of the stroke lesion. With computer vision and image processing techniques, we could automatically identify the lesion area instead of manually segmenting MRIs, which is time consuming. Previous work applied statistical machine learning methods, such as SVM [Lao Z., 2008] and Random Forest [Mitra J., 2014], to solve the problem, but suffer from complicated parameter tuning. Recently, deep learning networks have shown promising results on many visual recognition tasks (e.g. ImageNet [Krizhevsky A., 2012], LeNet [LeCun Y., 2013]). However, most of the methods are restricted to traditional 2D images, which are not suitable for the 3D MRI images. In this paper, we design an automatic lesion segmentation pipeline named Deep Lesion ConvNet segmentation pipeline by utilizing a 3D convolutional neural network (CNN) for the medical image modality. Our method can produce the stroke lesion mask without requiring parameter or threshold adjustments. The CNN is based on 3D anatomic T1 images but could incorporate additional MR contrasts. The high average dice coefficient achieved on our dataset demonstrates the effectiveness of our method.