Experimental MIP applications

 ^{Back to main research page:} 

Both to specify needs in PS and to validate my solutions, I study two applications, based on different approaches of MIP:
factor analysis programs, to identify functional information in temporal medical image sequences, based on their statistical properties;
computer vision programs, to perform brain segmentation, based on spatial characteristics (e.g. anatomic) of  images.

The construction of knowledge bases for these applications helps better identify the nature of the knowledge involved in the use of MIP, to design an adapted PS engine and to easy the expression of this knowledge.
 

Factor analysis for functional MIP

In collaboration with the INSERM Unit 494 (Quantitative medical imaging, directed by Pr. Todd-Pokropek), Paris, France, my first study concerns the Factor Analysis of Medical Image Sequences (FAMIS) method, used to estimate physiological functions that underly NM/MRI dynamic sequences (see article for SPIE Medical Imaging'97 conference). FAMIS is a typical example of a powerful MIP method, but complex and sparsely distributed: it solves different clinical goals based on performing statistical data analysis methods, involving many choices and parameters difficult to tune.
 
I have developed a knowledge base for FAMIS using the YAKL language. The reference application of FAMIS in that case is osteosarcoma chemotherapy follow-up, to predict the efficacy of the treatment.
 

Dynamic sequence of MRI images and results of FAMIS applied to estimate physiological functions that underly a knee osteosarcoma. Two first images present several sagittal views of tumoral knee, from initial sequence; third image shows the 3 extracted factors superimposed to their spatial location, respectively: rapid vascular factor, slow vascular factor and accumulation factor in soft tissues.

FAMIS sometimes requires the definition of one or more regions of interest (ROI) on input image sequence: their appropriateness and characteristics are determined thanks to expertise about the impact of focusing data on FAMIS results.
The choice of the best way to perform FAMIS aggregation substep of data preparation is submitted to knowledge about the way processing is influenced by the anatomical properties of the area to analyse (e.g. organ shape, vascularisation topology).
Furthermore, if geometric aggregation is chosen for instance, the shape and size of the geometric pattern it uses depends on methodological information about e.g. the quality of the image sequence, the shape and relative size of the area of interest in the image.
When it comes to evaluating results, expertise required is twofold: a rough idea of the physiological activity of the analysed area and its statistical behaviour through FAMIS, together with know-how about what can provoke FAMIS to issue bad results (such as movement in initial image sequence).
Whenever result quality is not satisfying, more expertise is necessary to decide how to correct FAMIS application.
 
 

(size: 33 Ko)

PEGASE engine has proven weak to take into account the flexibility needed to handle FAMIS use. This application is my main source of specifications and experiments for the MedIA engine.
An example scenario of FAMIS supervision with MedIA will soon be available. 

 
An extension of this knowledge base is in progress, in collaboration  with CERMEP-CREATIS, within the Hospital for Neuro-cardiology,  Lyon, France. In that case, FAMIS is used to characterise myocardic perfusion on dynamic sequences of PET or MRI images.

 
Computer vision for anatomic MIP

My second study concerns an application of brain anatomic segmentation on 3D MRI images based on computer vision methods (mainly mathematical morphology), in collaboration with Gregoire Malandain, from the EPIDAURE team, INRIA Sophia Antipolis. In that case, medical knowledge used by the expert is not included in programs but rather involved in expert's mental process. This study enables to define to what extent, in what form, and at what abstraction level medical knowledge is required to supervise such programs. Another interest in this application comes from expert's highly trial and error-based reasoning , as program parameter values are difficult to determine appropriately  in one go. Indeed, these are very sensitive to the acquisition quality of images to process. The repair phase of PS takes here its whole importance, and extends to the necessary storage of the history of choices made for processing. Decisions also consider the global objective of the processing, e.g. to determine with which precision brain must be extracted.

Example of a brain segmentation on 3D MRI images: first image is part of initial MRI brain acquisition, second image shows a slice of the result of segmentation and third image presents a superposition of the segmented area on initial image, for better visualisation. More about the method itself can be found on Gregoire Malandain's home page.