Considerations in Medical Problem Solving � What is already known by rrboy

VIEWS: 4 PAGES: 23

• pg 1
Considerations in Medical Problem Solving

• What is the form of the knowledge?

• How is the knowledge used?

• Does the knowledge consist of small independent chunks or is it large
and complex?

1. Gather clinical and physical data

2. Decide general approach

4. Use computer simulation to conﬁgure radiation beams

5. Verify feasibility

Repeat steps as necessary. . .
Deﬁning the GTV, CTV and PTV
Planning Target Volume Expansion Components

The Planning Target Volume allows for:

• tumor motion (internal)

• patient motion (external)

• variation in treatment setup (from treatment to treatment)
Determining the Planning Target Volume

Factors affecting target volume:

• Tumor site (variable granularity)

• Stage (primarily T-stage)

• Cell type, e.g., squamous cell, adenocarcinoma, etc.

Planning Target Volume Tool (PTVT)

Takes a tumor volume and expands it into a planning target volume.

1. Determine all the margin factors

2. Combine in a root-mean-square sum

3. Expand each (2-D) tumor contour radially from a “center of gravity”

4. Add a contour at each “end” to expand in the z direction
Implementation:

• Common LISP, CLOS

• Margin factors represented by production rules

• No user interaction - intended to be embedded in an RTP system
Example PTV Rules

(<- (setup-error ?x (0.8 0.8 0.8))   ;; (0.8,?,?) Verhey82,
;; approved SH 3/5/92
(immob-type none)))

(<- (pt-movement ?x (0.3 0.3 0.3)) ;; MAS/JMU-2/3/94
(immob-type none)))

(<- (setup-error ?x (0.5 0.5 0.5))   ;; SH 3/5/92

(<- (pt-movement ?x (0.1 0.1 0.1)) ;; MAS/JMU-2/3/94
The PTV Tool in Action

Prism demo!

• Manual interactive design

• Linear programming, simulated annealing (Rosen, et. al.)

• Inverse methods (Altschuler, Brahme, et. al)

• Heuristic search

Hypothesis:

1. Complex treatment designs can be found that are superior to typical
clinical practice.

2. An artiﬁcial intelligence program can ﬁnd such treatment plans faster
than human or numerical techniques.
Generate-and-test plan optimization

Generate Prototype Plans

Compute Dose Distribution

Find high and low dose regions

Generate           Compare with
"repaired" plans       other plans

most promising          "evaluated"
plans                 plan list
Prototype plans

• PO-LAT - parallel opposed ﬁelds, each compensating for the other, to
give uniform dose

• WEDGED-PAIR - treats a corner rather than a central region

• WEDGED-OBLIQUE - like WEDGED-PAIR but avoids a critical struc-
ture

• TRI-BEAM - treats a central target with obstacles
Treatment Plan Revision Rules

If: cold spot in deep area and hot spot in shallow area
Then: switch to higher energy

If: cold spot just outside a beam edge
Then: increase beam opening at that edge

If: organ overdosed by one beam
Then:

1. reduce beam weight, increase others

2. change gantry angle

3. add block to that beam
Coding Rules in Lisp

(<- (solution-p ?beam increase energy)
(and (type-p ?beam beam)
(penetrates-p ?beam shallow)))

(<- (penetrates-p ?beam shallow)
(and (type-p ?beam beam)
(type-p ?spot trouble-spot)
(dose-p ?spot cold)
(deepin-p ?spot ?beam)))

(<- (deepin-p ?spot ?beam)
(and (type-p ?spot trouble-spot)
(type-p ?beam beam)
(within-p ?spot ?beam)
(region-p ?spot deep)))
Managing the search for good plans

• Objective Evaluation describes a plan’s deﬁciencies

• Subjective Evaluation compares a plan to other plans

• A Similarity Metric is used to associate a plan with a region or cluster
of evaluated plans

• A Prioritizing Algorithm is used to select the most promising new plans
Objective evaluation:

• is a list of trouble spots

• used to generate revised (improved) plans

• also used for subjective evaluation

Subjective evaluation

• is a summary of pairwise comparision with other plans

• uses numerical properties of trouble spots

• is sometimes equivocal

• is used to characterize regions of the problem space

A Similarity Metric:

• associates a plan with a cluster of evaluated plans

• is used to prune the search for good plans
How did it turn out?

• Implemented only in two dimensions

• Initial cases showed that pruning improved the search efﬁciency by a factor of two.

• References:
´
– Paluszynski, W., Kalet, I. Design Optimization Using Dynamic Evaluation. Pro-
ceedings of the Eleventh International Joint Conference on Artiﬁcial Intelligence,
August 1989, Detroit, Michigan, American Association for Artiﬁcial Intelligence,
1989.
´
– Paluszynski, W. Designing radiation therapy for cancer, an approach to knowl-
edge based optimization. PhD thesis, University of Washington, 1990.
Medical Guidelines and Systems

Some history:

• Rule based “expert systems” for medical diagnosis and therapy

• Interoperability of rules - knowledge sharing strategies

• Arden Syntax - a standard text based interchange format

• GLIF - ﬂowchart based model, superset of Arden

• ASBRU - incorporates time and intention
Semantics of Guidelines

A guideline represents a single modular decision

• Is there such a thing in medicine?

• What input is needed (the “curley braces” problem)

• What does it mean to execute a guideline?
Arden Syntax

• Medical Logic Modules (MLM) represented in ASCII text ﬁles

• Procedural in style, but “frame-based”

• Can “call” each other, but not usually written that way
Conclusions

• Medical systems introduce engineering complications

• Medical knowledge is not categorical (need probability and defaults)

• Much remains to be done. . .

To top