Computer-aided detection of early esophageal cancer

Fons van der Sommen

1

Barrett’s cancer

2

Esophagus

Barrett’s tissueDiaphragm

Stomach Stomach

Diaphragm

Esophagus

Over 30 times increased risk of esophageal cancer!

Gastro-esophageal junction

Hiatal hernia

Treatment: late detection

3

Treatment: early detection = very hard…

http://www.ooso.org.ukhttp://surgicalgastroenterologist.com 4

If my phone can detect faces…

5

Could my endoscopedetect early cancer?

Goal: computer-aided detection system

Can we develop a system that helps the gastroenterologist find early Barrett’s cancer?

6

How do experts find these early cancers?From literature:

• “… the endoscope should be gradually withdrawn to examine the inflated Barrett’s segment in overview for any mucosal irregularities”

• “The primary step in diagnosis is to identify the presence of an area of the mucosa slightly discolored (more pale or more red), an irregular microvascular network, or a slight elevation or depression…”

7 8

Quantify discriminative information• Abnormal color

• Irregular mucosal patterns

9 10

Spectral analysis

van der Sommen et al., Neurocomputing 2014

11

123 85 176 10 12 22 85 43 92 30 75 84 10 33 8 3 2 7 6 4 5 4

Averagecolor

Std. dev. color

Averagetexture

Std. dev.texture

Each image patch quantified by a set of numbers:

Quantify local image information

van der Sommen et al., Neurocomputing 2014

Machine learning

12

Model

Mapping learned from examples

Machine learning

13

Model

Number of classes

Features Class confidence

(Classification)

+ Categoricale.g. {“Cancer”,

“No cancer”}

How good is our classification model?

14

15

ExpertsModel

van der Sommen et al., Endoscopy 2016 16

System Annotation

van der Sommen et al., Endoscopy 2016

17

System Annotation

van der Sommen et al., Endoscopy 2016

However, what is not captured by the camera,

cannot be detected…

18

19

Volumetric Laser Endomicroscopy

20

Courtesy prof. dr. Jacques BergmanAMC Amsterdam

21

Courtesy prof. dr. Jacques Bergman AMC Amsterdam

22

Courtesy prof. dr. Jacques Bergman AMC Amsterdam

Clinical prediction model (Swager et al., Gastrointestinal Endoscopy 2016)

23

AUC = 0.81

24

Quantify information: Layer Histogram

24

d

d

d

8

15

28

21

13

7

4

4

8

15

28

21

13

7

4

4

8

15

28

21

13

7

4

4

Feature vector

Intensity histograms

van der Sommen et al., SPIE Medical Imaging 2017

25 26

2 VLE experts Algorithm +/-

Sensitivity 0.83 0.90 + 8%

Specificity 0.71 0.93 + 31%

AUC 0.81 0.95 + 17%

Classification time 1-2 hours 90 ms - 99.998%(40.000 times faster)

Clinical validation on 60 VLE images with known histopathology

van der Sommen et al., IEEE Trans. Med. Im. (submitted)Swager et al., Gastrointest. Endosc. (submitted)

ConclusionsAs the amount of medical data and the signal complexity increase, computer vision methods can be of great help to medical doctors.I. A computer might recognize patterns which appear to be noise

to the human eye;

II. Algorithms are generally much faster in signal interpretation and can learn from millions of examples;

III. Machine learning algorithms can handle many signals at once, enabling large-scale, multi-modal analysis.

In addition, we found that using prior clinical knowledge typically leads to the best results, especially when not much data is available.

27 28

Prof. Jacques Bergman, MD, PhDFull Prof. Interventional Endoscopy,

Gastroenterologist, Oncologist

Anne-Fré Swager, MDPhD Candidate,

Interventional Endoscopy

Wouter Curvers, MD, PhDGastroenterologist,

Oncologist

Erik Schoon, MD, PhDDirector of Endoscopy,

Gastroenterologist, Oncologist

Fons van der SommenPhD Candidate

Medical Image Analysis

Svitlana Zinger, PhDAssistant professor

Medical Image Analysis

Prof. Peter de With, PhDFull professor

Video Content Analysis

Thank you on behalf of all partners!