Segmentation of ischemic stroke lesion from 3d mr images using random forest

This paper focuses on segmentation of ischemic stroke lesion from the dataset contributed by Ischemic Stroke Lesion Segmentation (ISLES)-2015 Sub-acute Ischemic Stroke lesion Segmentation (SISS) challenge. The dataset comprises 28 stroke cases of magnetic resonance (MR) images. This study considers fluid attenuation inversion recovery (FLAIR) MR images for the segmentation of lesions. MR images are affected by various artifacts and noise. Hence, we applied wavelet based data denoising technique by optimal parameter selection for stroke lesion enhancement followed by random forest (RF). RF classifier was trained for different part of the brain for segmenting the stroke lesion. The obtained results show best overall segmentation accuracy when compared with the other methods. To measure the image similarity between the ground truth and the segmented image, we used the evaluation method provided by the ISLES-2015. The average symmetric surface distance (ASSD) of the segmentation was measured to be 4.57 mm, while Dice's coefficient ((DC) lies between 0 and 1) was used to measure the volume overlap accuracy with an average of 0.67. The maximum of all surface distance was given by Hausdorff distance (HD) with an average of 28.09 mm. The average precision and average recall observed was 0.70 and 0.71 respectively. The ISLES image data and ground truth images kept on being openly accessible through an online virtual skeleton database which is an ongoing benchmarking resource.