Automatic white matter lesions detection and segmentation of brain magnetic resonance images

White matter lesions (WML) are frequently associated with neuronal degeneration in ageing and can be an important indicator of stroke, multiple sclerosis, dementia and other brain-related disorders. WML can be readily detected on Magnetic Resonance Imaging (MRI), but manual delineation of lesions by neuroradiologists is a time consuming and laborious task. Furthermore, MRI intensity scales are not standardised and do not have tissue-specific interpretation, leading to WML quantification inaccuracies and difficulties in interpreting their pathological relevance. Numerous studies have shown tremendous advances in WML segmentation, but flow artefact, image noise, incomplete skull stripping and inaccurate WML classification continue to yield False Positives (FP) that have limited the reliability and clinical utility of these approaches. The present study proposed a new MRI intensity standardisation and clustered texture feature method based on the K-means clustering algorithm. Enhanced clustered texture features and histogram features were constructed based on the proposed standardisation method to significantly reduce FP through a Random Forest algorithm. Subsequently, a local outlier identification method further refined the boundary of WML for the final segmentation. The method was validated with a test set of 32 scans (279 images), with a significant correlation coefficient (R=0.99574, p-value < 0.001) between the proposed method and manual delineation by a neuroradiologist. Furthermore, comparison against three state-of-the-art methods for the 32 scans demonstrated that the proposed method outperformed five of seven well-known evaluation metrics. This improved specificity in WML segmentation may thus improve the quantification of clinical WML burden to assess for correlations between WML load and distribution with neurodenegerative disease