Comparison of deep learning and conventional machine learning methods for classification of colon polyp types

Refika Sultan Doğan; Bülent Yılmaz

Uneingeschränkter Zugang

Comparison of deep learning and conventional machine learning methods for classification of colon polyp types

Refika Sultan Doğan

und

Bülent Yılmaz

| 22. Jan. 2021

The EuroBiotech Journal

Band 5 (2021): Heft 1 (January 2021)

Über diesen Artikel

Vorheriger Artikel

Nächster Artikel

Zitieren

Article Category: Research Article

Online veröffentlicht: 22. Jan. 2021

Seitenbereich: 34 - 42

DOI: https://doi.org/10.2478/ebtj-2021-0006

© 2021 Refika Sultan Doğan, Bülent Yılmaz, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Sample images obtained using WLI and NBI imaging modality during colonoscopy for different types of polyps: a) hyperplastic (WLI), b) hyperplastic (NBI), c) serrated (WLI), d) serrated (NBI), e) adenomatous (WLI), and f) adenomatous (NBI).

(a) Original colonoscopy image and (b) gray scale image.

Blue squares define a cell with 10x10 pixels yielding a total of 400 cells (a). On each cell HOG features were overlayed (b).

The number of test and training samples according to polyp-based stratification (80% polyps used in the training and 20% in the test set).

	Adenoma	Serrated	Hyperplastic
	Resection		No-resection
Training	32	12	17
Test	8	3	4

According to the polyps-based stratification, number of frames for each class (N: No-resection, R: Resection, A: Adenoma, H: Hyperplastic, S: Serrated).

	2-Class NBI		2-Class WLI		3-Class NBI			3-Class WLI
	N	R	N	R	A	H	S	A	H	S
Test	1162	4772	574	2769	3544	1162	1220	1990	574	779
Train	6991	22175	6795	26147	20504	6991	5964	21311	6795	4816

Two- and three-category classification results for different imaging modalities using deep learning.

	3NBI	3WLI	2NBI	2WLI
Accuracy	0.694	0.759	0.752	0.745
Recall	0.841	0.868	0.849	0.826
Precision	0.518	0.690	0.849	0.860
f-measure	0.517	0.726	0.849	0.843

Accuracy of the doctors’ predictions.

	A-H	A-H-S
Expert 1	0.82	0.64
Expert 2	0.83	0.69
Expert 3	0.78	0.65
Expert 4	0.77	0.58
Novice 1	0.78	0.60
Novice 2	0.86	0.68
Novice 3	0.75	0.51

Two- and three-category classification results for different imaging modalities using conventional machine learning.

	3NBI	3WLI	2NBI	2WLI
Accuracy	0.632	0.587	0.874	0.944
Recall	0.504	0.556	0.910	0.960
Precision	0.463	0.572	0.662	0.807
f-measure	0.483	0.564	0.767	0.877

Number of extracted frames for each class.

Class Types
	Adenoma	Serrated	Hyperplastic
Imaging Modality	Resection		No-resection
NBI	3544	1228	1162
WLI	1990	779	574

Accuracy of classification results.

Imaging Modality	Tissue Types	Machine Learning	Deep Learning
NBI	A-H	0.874	0.752
NBI	A-H-S	0.632	0.694
WLI	A-H	0.944	0.745
WLI	A-H-S	0.587	0.759

eISSN:: 2564-615X
Sprache:: Englisch

Zeitrahmen der Veröffentlichung:: 4 Hefte pro Jahr
Fachgebiete der Zeitschrift:: Biologie, Genetik, Biotechnologie, Bioinformatik, andere

Zeitschrift RSS Feed

Comparison of deep learning and conventional machine learning methods for classification of colon polyp types

Article Category: Research Article

Online veröffentlicht: 22. Jan. 2021

Seitenbereich: 34 - 42

DOI: https://doi.org/10.2478/ebtj-2021-0006

© 2021 Refika Sultan Doğan, Bülent Yılmaz, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Figure 1

Figure 2

Figure 3

Figure 4

The number of test and training samples according to polyp-based stratification (80% polyps used in the training and 20% in the test set).

According to the polyps-based stratification, number of frames for each class (N: No-resection, R: Resection, A: Adenoma, H: Hyperplastic, S: Serrated).

Two- and three-category classification results for different imaging modalities using deep learning.

Accuracy of the doctors’ predictions.

Two- and three-category classification results for different imaging modalities using conventional machine learning.

Number of extracted frames for each class.

Accuracy of classification results.