Understanding Brain Magnetic Resonance Images from Automatically Generated Interval-Valued Membership Functions

Medical images are representations of tissues and parts of the human body which play a crucial role in diagnosis assistance and human anatomy examination. New processing requirements related to problems of classification or segmentation are unceasingly generated. In this context, methods for segmentation that enables interpretable knowledge discovery can lead to significant contributions to the study and solution of certain medical problems. In a previous work, we proposed a data classification method called Type-2 Label-based Fuzzy Predicate Classification (T2-LFPC) which automatically generates interval-valued membership functions and predicates. In the present work, a methodology for interpreting brain magnetic resonance images in sequences PD, T1, and T2 with different levels of additive noise is proposed. Three measures on interval-valued membership functions are proposed and analyzed. Both simulated and real images are considered. The segmentation performance is consistent with the obtained with the test methods. The major contributions are a) the definition of attributes on the features and the association of them to each tissue, b) the description of relationships between attributes and tissues providing linguistic interpretation, c) the identification, quantification and description of both vagueness associated with the attributes and spread of intensities of pixels belonging to each tissue. Nevertheless, the knowledge achieved is consistent with what is known in the field of brain magnetic resonance images, which indicates that the methodology proposed constitutes a sound approach for knowledge discovery. Therefore, it could be extended to other medical imaging domains, making it a general approach for understanding medical images.