Detection and segmentation of large structures in an image or within a region of interest have received great attention in the medical image processing domains. However, the problem of small pathology detection and segmentation still remains an unresolved challenge due to the small size of these pathologies, their low contrast and variable position, shape and texture. In many contexts, early detection of these pathologies is critical in diagnosis and assessing the outcome of treatment. In this paper, we propose a probabilistic Adaptive Multi-level Conditional Random Fields (AMCRF) with the incorporation of higher order cliques for detecting and segmenting such pathologies. In the first level of our graphical model, a voxel-based CRF is used to identify candidate lesions. In the second level, in order to further remove falsely detected regions, a new CRF is developed that incorporates higher order textural features, which are invariant to rotation and local intensity distortions. At this level, higher order textures are considered together with the voxel-wise cliques to refine boundaries and is therefore adaptive. The proposed algorithm is tested in the context of detecting enhancing Multiple Sclerosis (MS) lesions in brain MRI, where the problem is further complicated as many of the enhancing voxels are associated with normal structures (i.e. blood vessels) or noise in the MRI. The algorithm is trained and tested on large multi-center clinical trials from Relapsing-Remitting MS patients. The effect of several different parameter learning and inference techniques is further investigated. When tested on 120 cases, the proposed method reaches a lesion detection rate of 90%, with very few false positive lesion counts on average, ranging from 0.17 for very small (3-5 voxels) to 0 for very large (50+ voxels) regions. The proposed model is further tested on a very large clinical trial containing 2770 scans where a high sensitivity of 91% with an average false positive count of 0.5 is achieved. Incorporation of contextual information at different scales is also explored. Finally, superior performance is shown upon comparing with Support Vector Machine (SVM), Random Forest and variant of an MRF.
Read full abstract