Approaching the complexity of biomedical signal processing

1
Approaching the complexity of biomedical signal
processing

• An agent-centered perspective
• Part I - Complexity
• Part II - Agent-centered design
• Part III - Application to patient monitoring

2
Complexity in biomedicine key issues

• (Biomedical signal) interpretation
• A task whose rationality is limited
• A task that is difficult to operationalize
• Keep the richness of the information at hand,
  avoid oversimplification or reductionism
• A systemic approach
• The situated agent paradigm

3
A task whose rationality is limited

• A capacity to discriminate and differenciate
  along with a capacity to associate and construct
  
• An infinite range of connexions but not all
  considered as relevant and significant at a time

4
The  unexpected  visitor

• Interior of a room with a group of people a
  piano in the background
• A man entering a room, he is wearing an overcoat
  and has a hat in his hand
• A woman is in foreground standing up from a chair
  looking towards a man entering the room
• A baby in a high chair, three other children in
  the background observing the visitor
• A woman in an apron by the door
• A.L.Yarbus, Eye Movements Vi-sion, Plentum
  Publish. Inc., 1967

 A text is an open universe where the interpret
may discover an infinite range of connexions -  
- U. Ecco, The limits of interpretation, 1990
5
Images as a universe of discourse

• The universe of images is contextually incomplete
  
• Taken in isolation, images have no assertive
  value but rely on some external context to
  predicate their content, and to endow them with
  meaning
• A single images, disconnected from any kind of
  external discourse, doesnt have any meaning that
  can be searched, unless we make some additional
  assumptions
• The image is explicitly linked to an external
  discourse, an intended message (eg annotated)
• The image is a priori inserted in a domain that
  is restricted enough so that one can disregard
  any other meanings (eg medicine, where images are
  interesting because of their diagnostic value)
• Santini 2002

6
Interpretation as a situated process

• A process that is context-sensitive a situation
  is perceived and make sense only in some context
  (neighbouring or past information, current
  hypotheses and goals)
• A process which do not obey any external
  predefined goal
• Rather an exploratory process according to which
  past perceptions give rise to further intentions
  driving further perceptions

7
A process that is context-sensitive

• A square perceived as light grey in the shadow vs
  a square perceived as dark grey outside the
  shadow their numerical values are the same!
• "Whilst part of what we perceive comes through
  our senses from the object before us, another
  part (and it may be the larger part) always comes
  out of our own mind."
• - W. James

8
The role of intentionnality

• Yarbus 67
• 1. No question asked
• 2. Judge economic status
• 3. Give the ages of the people
• 4. What were they doing before the visitor
  arrived ?
• 5. What clothes are they wearing ?
• 6. Remember the position of people and objects
• 7. How long is it since the visitor has seen the
  family ?

9
Scene understanding in medicine a situated
process

• The medical  scene  set of data and
  information that can be collected in the patient
  environment to support the diagnosis process
• The specificity of the medical diagnosis process
  lies in the capacity to evocate a restricted set
  of diagnosis hypotheses, based on a restricted
  set of queries, tests or investigations, these
  activities being grounded in a collection of
  environmental and patient-dependent factors
• Medical decision-making is anticipatory
  medical experts are known for their ability to
  rapidly sketch a situation, and then to select a
  small set of relevant hypotheses to constrain
  further analysis

10
Scene understanding in medicine a situated
process
11
The role of attentionnality

• From Daniel J. Simons 2003 - Surprising studies
  of visual awareness - Visual Cognition Lab. -
  University of Illinois
• http//viscog.beckman.uiuc.edu/djs_lab/

12
Scene understanding  we do not just see, we
look 

• The previous display examplifies our capacity to
  work in presence of a large amount of
  information, part of it is imperfect and noisy,
  part of it is purely not relevant to the task at
  hand. So we have to derive an approach which
  takes (more or less) explicitely into account
  this capacity
• There is no ideal representation, rather the
  necessity to constrain interpretation processes
  in an active way
• Diagnosis is the problem of controlling trial and
  error processes
• Diagnosis is an act of attention we do not
  just see, we look
• We act toward the external world in order to
  achieve a particular goal. There is no necessity
  to reconstruct everything, if it is not needed
  rather keep focused on what is needed with
  respect to the goal

13
Postulate

• Interpretation is not a representation activity
  
• Rather it is an exploratory activity guided by
  the search for new information, constraints and
  knowledge
• The purpose is not to exploit a priori models of
  the data, processings or goals,
• But rather to rely on interaction to dynamically
  acquire new knowledge elements, adapt to the
  changing environmental conditions and build
  solutions that are progressively more adapted to
  the potentially evolving context of the problem
  at hand

14
Consequences

• The role of the computerized system is no more to
  accumulate a large quantity of information, but
  rather to look for relevant and significant
  information
• Its primary objective is to question the
  environment and not to try and represent it
• The computerized system is then considered from
  the viewpoint of its capacity to  navigate 
  among a universe of informations, models, tools
  and strategies
• Reaching a state in the decision space generates
  the ability to look forward

15
Interpretation as an exploratory activity
16
Biomedical scene interpretation a task that is
difficult to operationalize

• Cope with
• The poor quality of data
• The poor efficiency of tools
• The lack of knowledge
• The lack of goals

17
The poor quality of data

• Medical data is sparse, voluminous, multimodal
  and time-dependent
• It is poorly reproducible, due to variations in
  the acquisition process also the same object
  may appear under several appearances
• It is incomplete, due to mistakes in the data
  collection / investigation protocol, or to
  missing, partially available or occluded
  information
• It is poorly informative, eg specific of an
  event, when taken in isolatio different objects
  may appear under similar appearances
• What is informative is the fact that a
  combination of information occurs in space and
  time

18
The poor efficiency of tools

• Processings are error-prone
• The quality of a given processing tool can hardly
  be predicted, since it depends on the properties
  of the object under interest and its context
• A priori knowledge of the situations to process
  is necessary to process them correctly
• The most efficient way to solve a problem is to
  already know how to solve it then one can avoid
  search entirely Minsky 86

è La vision un problème complexe
19
The poor efficiency of tools

• Linked to the lack of knowledge and to the lack
  of goal !
• A contour low-level event, frontier bet-ween 2
  homoge-neous regions or border of an object?
• A number of abs-traction levels, com-petences and
  goals

20
what the computer sees
21
The lack of knowledge

• Knowledge is limited
• To interpret a situation implies to know about
  how to associate situations to decisions - this
  is often provided in terms of expert know-how
• However, the expert knowledge is either to
  general, and does not cope with the large amount
  of specific cases to handle, or to specific, and
  non tractable
• Learning is moreover known as a combinatorial
  problem generate general models of all possible
  situations, together with all possible processing
  sequences
• As knowledge is limited, try to minimize the a
  priori by characterizing the situations at hand
•  Take the world as its own model  Brooks 91

22
The lack of goals

• Goals are ill-defined, and there is no
   universal  ground truth, provided from the
  outside the main task of any interpretation
  system is precisely to build a description of the
  environment in which it has to evolve
• While image descriptions measure precise image
  data, detached from their semantic content, user
  goals are cued in their semantic content
  (universe of discourse), but detached from their
  quantitative description
• Existence of a  semantic gap  , i.e. a  lack
  of coincidence between the information that a
  computerized system can extract from the data and
  the interpretation that the same data have for a
  user in a given situation  Smeulder et al,
  2000

23
A systemic approach
24
System design a recipe

• Open the information space
• 1. multiply the representations and processing
  styles keep the variety to avoid reductionism
  
• 2. collect and keep available as much information
  as possible on the situations at hand  
• 3. fuse information from various sources to
  render the interpretation context-specific

25
System design a recipe

• Increase the efficiency of tools
• 4. make inferences more local, but based on
  richer descriptions
• 5. reduce the scope of processing, spatially and
  semantically, in order to break down and
  specialize the tasks, thus reducing the semantic
  gap
• 6. use information as active constraints to drive
  low level processes
• 7. increase processing  utility  by applying
  them when and where relevant

26
System design a recipe

• Keep the dependencies
• 8. work in a situated way keep in mind the
  mutual dependency between the situation at hand,
  the applicable tools and the goal
• 9. progress more slowly, but in a more robust way
  base each step on accurate knowledge, in order
  that accurate knowledge be produced

27
A systemic approach

• The richness of the interpretation process
  finally depends on the capacity to confront,
  break and combine information obtained at
  different levels, providing a distributed,
  cooperative status to the interpretation task
• Information from all possible sources have to be
  considered to face the lack of constraints
• Interpretation can not be reduced to a process
  working in isolation, in a linear and univocal
  way
• Rather it results from the interaction between
  mutually dependent focusing and operating
  processes, working from different viewpoints at
  different abstraction levels
• This leads to a systemic paradigm

28
The situated agent paradigm

• The agents are situated physically (at a given
  spatial or temporal location), semantically (for
  a given goal or task) and functionnally (with
  given models or competences)

Goal Space
Agents
Knowledge Space
Model Space
Information Space
29
The situated agent paradigm

• Situated agents
• Agents being anchored at a given position in the
  problem space, in terms of data to analyze, goals
  to be pursued and models to proceed
• These agents work in a specialized and local way,
  they produce partial results that are shared via
  the environment
• A dual adaptation
• Internal adaptation by the selection of adequate
  processing models, according to the situations to
  be faced and to the goals to be reached
• External adaptation by the dynamical generation
  of constraints, eg of new sets of data and goals
  to explore such adaptation may requires the
  creation of new agents, modifying as a
  consequence the structure of the analyzing system
  itself

30
The situated agent paradigm

• As the system works, it
• increase its confidence è more robust results
• completes its exploration è more complete
  results
• accumulates information è more adapted behavior

Goals
Knowledge
Models
Information
31
A case example
32
A case example

• Two mutually dependent processes
• Contour following triggered at successive steps
  of the region growing process limit their
  expansion
• Region growing triggered in case of failure of
  the contour following provide refined
  contextual information
• Each process works locally and incrementally,
  under dynamically and mutually elaborated
  constraints

33
A case example

• Successive focusings
• Process localization and state
• executing
• active
• waiting
• Process linkage
• seed process
• System load
• Segmentation result

34
A case example

• An Evolving Processing Structure
• A coupling between
• A dynamically evolving processing structure
• A dynamically evolving description of the initial
  image
• Towards an Agent-Centered Design
• A paradigm that steps back from classical
  procedural design
• A processing approach where the time, content and
  partners of the interaction are not planned in
  advance
• A processing approach which is not thought out in
  terms of chains or linking but in terms of
  interaction
• A problem solving approach where the solution is
  not sought in a global way