skin diagnosis

Question 1

what is the purpose of skin photography?

Answer 1

documentation of the localization, extent, surface, degree of inflammation, … of the skin.

This data can easily be compared

Question 2

epicutaneous test: what important decision has to be made? How can we differentiate the reactions?

Answer 2

allergic VS irritant.

Allergic reactions cause overheating -> infrared reading

Question 3

machine learning: three steps in the process of inference

Answer 3

1) observe a phenomenon
2) construct a model of the phenomenon
3) do predictions

Question 4

What are the three main frameworks in machine learning?

Answer 4

supervised
unsupervised
semi-supervised

Question 5

what is one risk with too complex machine learning models?

Answer 5

overfitting the data

Question 6

how many images do we have to use when traning a model (for example differentiate acne vs rosacea)

Answer 6

thousands of images -> be in the accuracy range of a medical application

Question 7

hand eczema: % of working population, what is key to rapid treatment, one problem with mannual annotations

Answer 7

• 10%
• early detection is key
• bias / low inter-annotator agreement

Question 8

segmentation: what are three things to take into consideration when evaluation the relation to the actual clinic?

Answer 8

• applicability: does it correlate with the surface values of the consultation
• computer vision and human vision: do they match
• human reproducibility

Question 9

how will AI be used in image diagnostics? will it be a substitute to human prediction?

Answer 9

no, it will be a decision support. There are too few images of rare diseases, human annotations are still needed

Question 10

can we train a model with no annotations? what does it require?

Answer 10

requires healthy images -> identifies abnormal areas, doesn’t need to compute backward step, robust to domain shifts

Question 11

what can we do to reduce noisy annotations?

Answer 11

use images from the same patient on training and test sets

Question 12

are current approaches data centric or model centric? what might solvethe bias in the models?

Answer 12

data-centric

generative networks

Question 13

acne: % of teenagers, how can we evaluate the treatment and the progression of the patient?

Answer 13

85%

daily measurement of the skin changes

Question 14

what can we use to look at the daily skin changes? for acne for example

Answer 14

3D facial reconstruction on cell phone -> immediate user feedback, data transfer to dermatologist for assessment

Question 15

what is the vision of modern dermatology?

Answer 15

• imaging on every visit
• comparison with findings / other patients
• recommendations for diagnostics and therapy

Question 16

what are 2 big advantages of teledermatology (based on photos that the patients take themselves / online appointements) ?

Answer 16

temporal evolution is represented !!

also able to help people from other countries (Africa), where resources are scarce but people have internet connection