Artificial Intelligence in Medicine: Why and How

1
Artificial Intelligence in Medicine Why and How?

• Guillermo Callahan Olachea.
• Helsinki University of Technology
• 68534B

2
Medicine and Artificial Intelligence

• Medicine, like science itself, is a huge
  cognitive process, but it is also to a great
  extent an Information management task, since
  decision making is based upon expert knowledge,
  information of the pacient and the doctors
  experience.

3
Medicine and Artificial Intelligence

• Since the knowledge, and its general
  improvements, in medicine are growing
  exponentially, and since it is also a branch that
  requires various points of view from data, is
  there another way to provide more knowledge to a
  physician so he can be able to make better
  decisions?

4
Medicine and Artificial Intelligence

• Medicine and Artificial Inteligence have been
  toguether for a little more than 10 years to be
  able to tackle various problems and enhance
  medical efficiency and quality.

5
Medicine and Artificial Intelligence

• Theres been many applications for Artificial
  Inteligence in Medicine. Not all of them have
  been as efficient and the applications, while
  tackling the same problem, can have a completely
  different idea and structure.

6
AI applications in Medicine

• Medical Education.
• Heart Condition or Disease Diagnosis.
• Breast Cancer Diagnosis.
• Breast Cancer Treatment.
• Acute Lymphoblastic Leukemia Diagnosis.
• Temporal Model Based Diagnosis.

7
AI in Medical Education

• The investigation in AI in Medical Education is
  based on the thought that AI has not lived up to
  its potential, since it provides the points
  mentioned in the next slide.

8
AI in Medical Education

• Applications do not save the personels time.
• Procedures for getting the needed desicion
  support take too long.
• Quality of decisión might not be satisfactory in
  some situations.
• Personel do not trust the applications.
• Personel see the applications as a threat.

9
AI in Medical Education

• Research on the topic mentions that by rising the
  education for medical graduates, using AI
  applications instead of providing solutions,
  provides solutions to rise the quality of medical
  procedures.
• This is by rising self-learning, self-esteem,
  self awareness, etc.
• Examples of this Informed, RadTutor. ILE-VT.

10
AI in Medical Education
11
AI in Heart Condition Diagnosis

• Theres some literature about AI related to the
  diagnosis of abnormal heart conditions. Since
  heart diseases are one of the most common deaths
  in the US.

12
AI in Heart Condition Diagnosis

• In the research literature, there was a finding
  of two applications.
• The implementation of a MLP based MDSS for heart
  disease diagnosis.
• WeAidU a MDSS for myocardial perfussion images
  using ANN.

13
MLP-Based MDSS for Heart Disease Diagnosis.

• Literature mentions since heart diseases can be a
  quite extensive and experience-required
  diagnosis, the implementation of a MDSS using
  soft computing can be effective for the detection
  of various heart diseases.
• This proposal uses a Multilayer Perceptron with
  an improved back propagation algorithm.

14
MLP-Based MDSS for Heart Disease Diagnosis.
15
MLP-Based MDSS for Heart Disease Diagnosis.

• The improved back propagation algorithm mentioned
  in the research includes
• A momentum term.
• An adaptive learning rate.
• A learning algorithm with forgeting mechanics.
• An optimized algorithm based on the conjugate
  gradients method.

16
MLP-Based MDSS for Heart Disease Diagnosis.

• The proposed system is a three layer MLP with 40
  input variables, 15 hidden nodes and 5 outputs.
  The learning algorith is explained on the
  following slide.

17
MLP-Based MDSS for Heart Disease Diagnosis.

• Testing by using the cross validation method,
  holdout test and the bootstrapping method
  indicates that the system is effective since it
  can accurately detect all five of the mentioned
  heart diseases (gt90) with comparable small
  intervals (lt05).

18
WeAidU a MDSS for myocardial perfussion images
using ANN.

• WeAidU is a computer based DSS system for the
  automated interpretation of diagnostic heart
  images. Which is available on the Internet
  (www.weaidu.com).
• The system is based on ANN, Image processing
  techniques and large well-validated medical
  databases.

19
WeAidU a MDSS for myocardial perfussion images
using ANN.
20
WeAidU a MDSS for myocardial perfussion images
using ANN.

• The DSS currently delivers two diagnostic advise,
  one regarding the precense of infarction and one
  that concerns ischemia. And the heart is divided
  in 5 parts and a diagnostic advise is given for
  each one of them.

21
WeAidU a MDSS for myocardial perfussion images
using ANN.

• The system uses 10 different ANN classifiers, one
  for each advise given and each classifier
  consists of an essemble of single ANN.
• The individual members of the essemble are single
  MLPs with a hidden layer of 5-15 nodes and one
  output.

22
WeAidU a MDSS for myocardial perfussion images
using ANN.

• For activation, the system uses the Tanh()
  function and each MLP is trained using gradient
  descent appied to a cross-entropy function. The
  gradicent descent method is augmented with a
  traditional momentum term and Langevin Extention.

23
WeAidU a MDSS for myocardial perfussion images
using ANN.

• The performance of the ANNs for detecting the
  diseases in different parts of the heart,
  measured as areas under the ROC curves, is in
  range of 83-96.
• This means that the tool has a very high
  potential for the tool as a MDSS.

24
AI in Breast Cancer Detection, Treatment and
Diagnosis

• Most of the medical investigations found for this
  presentation were in some form related to Breast
  cancer.
• On the research encountered there were three
  mayor contributions
• OncoDoc Computer Based Guideline for Breast
  Cancer Treatment.
• Research on selection for a MDSS for Breast
  Cancer Detection.
• A MOE MDSS for Breast Cancer Detection.

25
Research on selection for a MDSS for Breast
Cancer Detection.

• Following research on Breast Cancer Detection
  consists of finding the most acurate ANN for
  cancer detection.
• One research consists of by using a SOM ANN, to
  be able to find

26
OncoDoc A Computer Based Guideline for
physicians.

• OncoDoc is a Knowledge system that is based on
  providing medical guidelines to Doctors to
  provide the best diagnosis.
• Its aim its to provide a physician controlled
  operation of guideline knowledge through
  hypertextual reading of a knowledge based encoded
  as a decisión tree.

27
OncoDoc A Computer Based Guideline for
physicians.
28
OncoDoc A Computer Based Guideline for
physicians.
29
OncoDoc A Computer Based Guideline for
physicians.
30
OncoDoc A Computer Based Guideline for
physicians.

• The results from using this system the
  Pitié-Salpêtrière Hospital. It was found that it
  solved one of its initial objectives. As
  physicians valued OncoDoc with 96.60 of
  adherence and 64.28 of compliance.

31
AI in Breast Cancer Detection, Treatment and
Diagnosis

• The following research consisted of encountering
  the best possible fit to a SOM ANN for the
  detection and diagnosis of Breast Cancer.
• The procedure consisted of creating a SOM with
  clinical data, and encounter that function that
  resembled the most effective.

32
AI in Breast Cancer Detection, Treatment and
Diagnosis
33
AI in Breast Cancer Detection, Treatment and
Diagnosis
34
AI in Breast Cancer Detection, Treatment and
Diagnosis

• After finding the most accurate and effective
  eccuation, the team went on to find the
  quantitative method that most suited the SOM
  function. If not, it went to non-parametric
  methods to ANN.
• The methods analysed were Logistic Regression,
  Linear Discriminant Analysis, MLP, MOE, GR, RBF.

35
AI in Breast Cancer Detection, Treatment and
Diagnosis
36
AI in Breast Cancer Detection, Treatment and
Diagnosis
37
AI in Breast Cancer Detection, Treatment and
Diagnosis
38
(No Transcript)
39
(No Transcript)
40
AI in Breast Cancer Detection, Treatment and
Diagnosis
41
AI in Breast Cancer Detection, Treatment and
Diagnosis

• As a conclution of this investigation, it was
  mentioned that SOM can be an effective model for
  MDSS selection.
• Linear Discriminant Analysis was chosen as the
  best model for this type of analysis.
• The development of a collection of stacked
  predictors can provide noticiable improvement in
  generalization hability for both cases.

42
A MOE MDSS for Breast Cancer Detection.

• After the last research was created, another
  researcher created his own investigation that
  placed the MOE ANN as the best solution to
  diagnose breast cancer.
• As a difference to the last research that
  included MOE. This one used the
  Expectation-Maximization algorithm.

43
A MOE MDSS for Breast Cancer Detection.

• Here, both the experts and the gating networks,
  were MLPs, this on the theory that MLPs have
  the abbility to learn, and generalize, smaller
  training set requirements, fast operation and a
  mayor ease of implementation.

44
A MOE MDSS for Breast Cancer Detection.
45
Fuzzy Networks In Medicine.

• Acute Lymphoblastic Leukemia Diagnosis.
• Temporal Model Based Diagnosis.

46
Conclusions
47
Discusion

• Sorry for taking so long and not going to
  anything specificlythere was just too much
  information and so little time.
• Thank You for your Attention and Patience.