WEBVTT - generated by Videoportal FH Dortmund

1
00:00:01.138 --> 00:00:04.877
Skin lesions are a part of our everyday lifes,
throughout all ages.

2
00:00:05.160 --> 00:00:10.658
Wounds are the most common example. A scratched
knee, a blister on the foot, a burnt finger...

3
00:00:10.658 --> 00:00:14.754
we all know these little accidents that usually
heal well.

4
00:00:16.680 --> 00:00:23.972
As a rule of thumb one can say: The larger
a wound, the longer it takes to heal sufficiently
and treatment gets more complex.

5
00:00:24.021 --> 00:00:31.320
Yet not only the size of a wound determines
its healing process, but also age and conditions.

6
00:00:31.320 --> 00:00:38.392
In unfortunate cases that are not that uncommon,
acquired wounds can turn chronic.

7
00:00:38.392 --> 00:00:43.239
These require extensive care and continuous
monitoring over months up to years

8
00:00:43.239 --> 00:00:48.600
to enable a sufficient healing – and boy,
is that tricky to accomplish!

9
00:00:49.741 --> 00:00:52.827
Okay, now let’s have some wound basics!

10
00:00:52.827 --> 00:00:59.520
A chronic wound does not only show flesh and
skin. Three main tissue types can be classified
in the wound bed:

11
00:00:59.520 --> 00:01:05.000
Granulation, slough, and necrosis. Granulation
rocks, the rest sucks.

12
00:01:05.000 --> 00:01:12.000
Why, you ask? Well, basically, as granulation
enables epithelialization, the closing of the
wound.

13
00:01:12.000 --> 00:01:18.000
On the other hand, slough disturbs the healing
process, and necrosis further facilitates possible
infections.

14
00:01:18.000 --> 00:01:27.000
Both need to be sufficiently debrided, creating
fresh wound bed areas. For these, healing has
to start once again.

15
00:01:29.000 --> 00:01:37.000
In standard wound care, nowadays documentation
is supported by taking pictures of individual
wounds to track their progression over time.

16
00:01:37.000 --> 00:01:40.000
Remember, healing can take months up to years.

17
00:01:40.000 --> 00:01:46.000
However, neither slowly manifesting improvement
or aggravation are traceable adequately over
time,

18
00:01:46.000 --> 00:01:51.000
nor are rather dynamic changes in tissue composition
and appearance comprehensible.

19
00:01:51.000 --> 00:01:57.000
This is due to the nature of wounds which appearance
can get very complex,

20
00:01:57.000 --> 00:02:07.000
taking into account their shape, depth, colors,
tissue mixes, wetness, sterility, blood supply,
and many other criteria.

21
00:02:09.000 --> 00:02:13.000
So, how to answer questions that are relevant
to provide optimal care?

22
00:02:13.000 --> 00:02:17.000
For instance: Where does a wound start and
where does it it end?

23
00:02:17.000 --> 00:02:23.000
Which areas belong to which tissue type? And
how did the wound change over time?

24
00:02:23.000 --> 00:02:30.000
These questions can not be answered appropriately
by humans as their perception and conclusions
are subjective by nature.

25
00:02:30.000 --> 00:02:38.000
Two caregivers may present two more or less
different anwers for each question, as their
expertise and experience varies.

26
00:02:40.000 --> 00:02:47.000
Machine learning methods, however, are objective.
Specific models allow to recognize and analyze
patterns in images.

27
00:02:47.000 --> 00:02:52.000
Applying such methods for wound recognition
and tissue classification has the potential

28
00:02:52.000 --> 00:02:56.000
to lower the burden of caregivers and increase
individual care.

29
00:02:56.000 --> 00:03:03.000
Yet, their performance highly is highly depending
on the data they are trained with.

30
00:03:03.000 --> 00:03:09.000
Achieving appropriate results is challenging
due to the mentioned high variability of wounds.

31
00:03:09.000 --> 00:03:14.000
Hence, to yield sufficiently performing models,
a vast amount of images with a high variability

32
00:03:14.000 --> 00:03:21.000
as well as correct and unambiguous formal descriptions,
so-called annotations, is needed.

33
00:03:22.000 --> 00:03:27.000
Unfortunately, free and ethically approved
wound image databases are rare.

34
00:03:27.000 --> 00:03:32.000
Their combined amount of images is limited,
and data incorporates several issues.

35
00:03:32.000 --> 00:03:37.000
For instance: Images highly vary in quality,
given annotations are even rarer...

36
00:03:37.000 --> 00:03:41.000
... and existing ones highly differ in quality
and consensus.

37
00:03:41.000 --> 00:03:45.000
Also some concepts are underrepresented.

38
00:03:45.000 --> 00:03:52.000
For instance, a very prominent problem discussed
in machine learning research is the lack of
image data with darker skin tones.

39
00:03:52.000 --> 00:03:59.000
The consequence is, that approaches trained
with lacking concepts may not perform sufficiently
for each and everyone.

40
00:03:59.000 --> 00:04:06.000
But there are ways to cope with these issues:
Generative Adversarial Networks.

41
00:04:06.000 --> 00:04:14.000
These are able to learn individual concepts
from images and generate synthetic high-quality
images via blueprints.

42
00:04:14.000 --> 00:04:20.000
The main goal of this PhD project is to establish
an iteratively improving machine learning workflow

43
00:04:20.000 --> 00:04:25.000
workflow for an exact and robust analysis of
different types of skin lesions.

44
00:04:25.000 --> 00:04:38.000
Lack of data and underrepresented concepts
will be compensated artificially to enable
balanced and unified mod-els with increased
detection and classification performance.

45
00:04:38.000 --> 00:04:47.000
A patient’s wound healing history will be
traceable and comprehensible, enabling precise
treatment adjustments when indicated.

46
00:04:47.000 --> 00:05:04.000
This can accelerate the wound healing process
as well as reduce the risk of complications
that could otherwise lead to prolonged treatment
or amputations.