Author Email: firstname.lastname@example.org
Deep Learning proves to be the best way towards image classification tasks replacing ensembles with handcrafted features. It also has applications in medical imaging such as classification of tumorous and non-tumorous part from rest of the MRI image. Lesions is one such disease occurring in brain due to lack of oxygen during stroke. Automatic segmentation of ischemic stroke lesions is the perfect problem solved by using deep learning methods especially Convolution Neural Networks (CNNs). We present a patch based approach using multi-scale convolution layers with three pathways to segment lesions using 6 modalities of MRI images. The accuracy achieved using this method is comparable to other methods.
Ischemic Stroke Lesions are caused by lack of oxygen to the brain during stroke and are major causes of disability and death globally. Eighty percent of strokes are caused due to cerebral ischemia. Various treatments currently available are surgery, chemotherapy, radiotherapy, or a combination of both of them. MRI is very useful in finding lesions in clinical practice, as we can acquire MRI sequences providing different information of brain tissues (S. Bauer et al., 2013). It is important to find the lesions in early stages so that good follow up treatment is given fast by doctors. However, segmenting these lesions (D. N. Louis et al., 2007) is time consuming task and requires highly qualified neurosurgeons. Thus, physicians usually use rough measures for evaluation. For these reasons, accurate semi-automatic or automatic methods are required (S. Bauer et al., 2013). However, it is a challenging task, since the shape, structure, and location of these abnormalities contain very high variability. Additionally, the tumor mass effect change the arrangement of the surrounding normal tissues (B. Menze et al., 2015). Also, MRI images present problems as intensity inhomogeneity, or different intensity ranges among the same sequences and acquisition scanners.
Various methods for brain tumor (lesions) segmentation tasks have been proposed and one method is using CNNs for this task as they have proven to best methods to handle such kind of tasks. These methods use labeled image data to extract features and use those for inference. Four different models for CNNs are compared in this review using various evaluation metrics such as dice score, sensitivity and specificity. These methods are for multi-class pixel wise classification segmentation and we want to achieve a binary segmentation using and extending these methods. A novel patch wise three-pathway fully convolutional architecture is proposed for ischemic stroke lesion segmentation task.
CNNs have drastically improved the task of image recognition, segmentation and tracking in image processing. The applications of these methods are helpful in variety of fields like object tracking, security, medical imaging etc. Cascaded CNNs especially have been helpful in medical imaging as these methods provide fast training and inference. Patches based method are good at imbalanced class segmentation problems but pose a lot of problems and requires a lot of fine tuning. Using patches increases the inference time by a very high margin. Thus the task of automatic lesion segmentation at practical level is still pending
419 total views, 1 views today