Interpreting magnetic resonance images by means of fuzzy memberships functions

GUSTAVO J. MESCHINO; VIRGINIA L. BALLARIN; DIEGO S. COMAS

Seúl

Congreso; International Conference on Biomedical and Health Informatics 2021 - ICBHI 2021; 2021

KOREAN SOCIETY OF MEDICAL AND BIOLOGICAL ENGINEERING

Medical imaging is currently one of the central resources for the evaluation and diagnosis of patholo-gies. Technological advance continuously generates new types of medical images, substantially increas-ing the available medical information and defining new requirements for its analysis. Considering digi-tal image processing applied to medical images, a segmentation or a classification technique that al-lows discovering interpretable knowledge becomes extremely important as it can provide new knowledge regarding the type of images or the problem, which can lead to significant contributions for studying and solving medical problems. In previous works, we proposed a fuzzy-predicate-based clas-sification method which was applied to the estimation of the volume of tissues in brain Magnetic Reso-nance Images (MRI) considering T1-sequences. Besides performing image segmentation, that method automatically generates membership functions from prototypes extracted from labelled examples (Gold-Standard) enabling knowledge discovery. In the present work, we propose a methodology for segmenting and interpreting brain MRI from automatically discovered fuzzy membership functions. The study is focused on both simulated and real MRI (with their corresponding Gold-Standard) in se-quences PD, T1, and T2. The segmentation results indicate estimated accuracies, by 10-fold cross-validation, of 0.988±0.016 and 0.848±0.018 respectively for simulated and real images, overcoming test methods. The methodology proposed for interpretation of the resulting membership functions and predicates allows: a) to define attributes on the features and to associate them to each tissue, b) to de-scribe relationships between attributes and tissues providing linguistic descriptions, c) to identify vagueness associated with the attributes. The results of this methodology indicate it is a sound ap-proach for knowledge discovering: considering the previous knowledge about brain MRI, including medical experts? opinions, membership functions, predicates, and attributes are consistent. The meth-odology can be extended to other medical imaging domains, making it a general approach for inter-preting medical images.