ARTIFICIAL INTELLIGENCE IN NEUROIMAGING
Florence GOMBERT UTC, France 22-04-04
OVERVIEW General presentation of AI AI in Medicine AI in Neuroimaging Example of Expert Systems in neuroimaging
INTRODUCTION AI goal : Concept systems able to reproduce the human behaviour in its reasoning activity Modeling intelligence as phenomenon (as physics, chemistry that model some others phenomenon) Study intelligence from 2 perspectives: analysis : human intelligent functions synthesis : describe intelligent functions
AI GENERALITIES AI main points : Turing test : Are machines able to think? Searle, Chinese room (1980) : just behaving intelligently isn't enough 2 schools : -creating computer systems whose behaviour is at same level indistinguishable from humans ( Strong AI ), - systems are not intended to have an independent existence (Weak AI ) ‘cognitive prosthesis’
AI methods compare to classical informatics methods Classical methods
AI methods
Close to machine functioning
Close to human functioning
Number or text processing
Symbols processing
Use a lot of calculation Follow straight and exhaustive algorithms Can't be generalize to only one class of pb
Use a lot of inferences Use heuristics and uncertain reasoning Can be generalize to different fields
Some main fields of AI Experts systems : medicine, financial analysis… Robotic and CFAO Language understanding and automatic translation Pattern recognition Learning
2 - AI in Medicine Dream to create an “electronic brain”. This search to create artificially intelligent (AI) computer systems one of the most ambitious and, not surprisingly, controversial. With intelligent computers able to store and process vast stores of knowledge, the hope was that they would become perfect ‘doctors in a box’, assisting or surpassing clinicians with tasks like diagnosis. Very quickly become an attractive field
2 - AI in Medicine In 1984, Clancey and Shortliffe provided the following definition: “ Medical artificial intelligence is primarily concerned with the construction of AI programs that perform diagnosis and make therapy recommendations. Unlike medical applications based on other programming methods, such as purely statistical and probabilistic methods, medical AI programs are based on symbolic models of disease entities and their relationship to patient factors and clinical manifestations.”
2 - AI in Medicine Today, AIM systems are more typically describe as clinical decision support systems (CDSS). CDSS require the existence of ERP Support medication prescribing, An AI system could be running within an electronic patient record system, alert a clinician - when it detects a contraindication to a planned treatment. - when it detected patterns in clinical data that suggested significant changes in a patient’s condition. Reason with clinical knowledge, AIM create and use it
2 - AI in Medicine KB Systems = Commonest type of CDSS in routine clinical use. Also know as Expert Systems Contain clinical knowledge
Different types of clinical task to which expert systems can be applied Alerts and reminders Diagnostic assistance. Therapy critiquing and planning Prescribing decision support systems Information retrieval. Image recognition and interpretation. interpretation
2 - AI in Medicine Some CDSS have capacity to learn, leading to the discovery of new phenoma and the creation of medical knowledge. These machine learning systems can be used to: - develop the knowledge bases used by expert systems, - assist in the design of new drugs, - advance research in the development of pathophysiological models from experimental data
2 - AI in Medicine In the first decade AIM, develop clinicians assistants in process of diagnosis : DX plain, clinical decision support systems, 1987
Assist the process of diagnosis, and produce a ranked list of diagnosis Provides justification for each of differential diagnosis, and
suggests further investigations
3 - AI in Neuroimaging
Neuroimaging = intersection image processing, computer science, physics, statistics, neurosciences. Challenge: intelligent data analysis, information fusion, archiving, validation and visualization Use of AI methods and techniques
2 complementary view of AI : - engineering discipline, creation of Intelligent Machine - empirical science computational modeling of human intelligence
PB: - how distributed systems of brain areas work altogether - how mental capabilities emerge from neuronal networks - how knowledge and its representation by the brain can be attacked in computational terms or information process terms
3 - AI in Neuroimaging
Relationships brain architecture / cognitive models.
Lack of architectural information in MR images Development of brain mapping strategies based on various coordinate systems without accurate architectural content. content
Great interest for structural strategies used in computer vision Key point = conversion of the raw images into structural representations before analysis.
3 - AI in Neuroimaging The Spatial Normalization paradigm :
Comparison and theorization of anatomic localization results
Inter-individual variability of brain shape
Need a previous pairing of functional images
Most try to make disappear the macroscopic variable shape
Try to have identical brain Spatial Normalization
3 - AI in Neuroimaging The Spatial Normalization paradigm :
Spatial normalization + proportionality factors (to reduce the variability) = Brain reference in 3D (Talairach referential)
3 - AI in Neuroimaging The Proportional Normalization paradigm (Atlas):
Relies on 3D coordinates indicating a location in a template, and is
called the proportional system.
3D indicate a localization in a reference brain
Reference : 3D deformation aligns and minimizes the individual
macroscopic anatomy with the template brain anatomy
Can compare functional image from different individual point by
point
3 - AI in Neuroimaging
3 - AI in Neuroimaging
Structural methods: basis « Group study » look for, in the total
individual activation, these
reproducible over people. Use anatomical localization of activation Very different pairing Connected group points are paired, not pixels Each group can represent a cognitive
module
3 - AI in Neuroimaging Structural methods: basis
Link to vision paradigm
Raw image is reduced to a
graph
Graph represent the organization of local information extracted
by filter
This representation is confronted to structural representation a
priori known
3 - AI in Neuroimaging Raw data, T1
Brain segmentation
RH external cortex
RH inner cortex
Skeleton of cortex + CSF
segmentation of skeleton
3 - AI in Neuroimaging Future of Structural methods in brain mapping A lot of structural approach of brain image relies on homogenous structural data Future goal : mix several kinds of data into a more complex structural representation For example, cortical folds and activation related blobs should be mixed to study the potential localization power of the sulco/gyral patterns. Attractive example consists in using MR diffusion data to infer the connectivity between activated clusters, or between cortical gyri.
3 - AI in Neuroimaging Future of Structural methods in brain mapping What we can expect, in the future : - Convert raw images into structural representations. - Merge several representations coming from the same individual. - Match each new structural representation with one syntactically corresponding structural model. This model could stem from a database of the current knowledge, which could include concurrent points of view. - Infer new similarities across a set of individuals to improve the current structural models.
3 - AI in Neuroimaging Future of Structural methods in brain mapping
Extract from : J.-F. Mangin et al. / Artificial Intelligence in Medicine 30 (2004) 177–197
4 – Example of expert system in neuroimaging : Perfex Automatic interpretation of Cardiac SPECT data Assist diagnosis of coronary artery disease Employ KB methods to process and map 3D visual information in combination with related clinical data Infer structure (anatomy) from function (physiology) Assess the extent and severity of cardiovascular disease, quantitatively and qualitatively Presents the resulting diagnostic recommendations in numerical textual and visual forms in an interactive framework, thereby enhancing overall utility
4 – Example of expert system in neuroimaging : Perfex
http://www.cc.gatech.edu/gvu/visualization/perfex/d emo/demo3
4 – Example of expert system in neuroimaging : Thallium Diagnostic Workstation Diagnose thallium myocardial scintigraphy from a training set of examples.
Digitized images acquired by a gamma camera are send to TDW via Ethernet.
TDW uses induction to learn diagnostic rules.
Users can view imagery, enter their own findings and diagnosis, select a rule from a catalogue of rule sets (each rule is accompanied by performance statistics), and perform automated diagnosis.
CONCLUSION
Benefits from CDSS : 3 broad categories (Sintchenko et al., 2002): -Improved patient safety -Improved quality of care -Improved efficiency in health
care delivery
Improve Structural methods in neuroimaging
Use multimodality and data from different level and scale for a better brain architectonic
Develop Functional connectivity, anatomic connectivity //new technics
Links Book: Guide to Health Informatics 2nd Edition, Enrico Coiera,
Main Papers : - Artificial intelligence in neuroimaging: four challenges to improve interpretation of brain images - Coordinate-based versus structural approaches to brain image analysis
Website, Perfex: http://www.cc.gatech.edu/gvu/visualization/perfex/
Thanks and Questions...
