Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images

Yung Han Chen; Yi Jen Ju; Juinn Dar Huang

doi:10.1007/978-3-031-26354-5_5

Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images

Yung Han Chen^*, Yi Jen Ju, Juinn Dar Huang

^*Corresponding author for this work

Institute of Electronics

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

1 Scopus citations

Abstract

For the Diabetic Foot Ulcer Challenge 2022 (DFUC2022) hosted by MICCAI 2022, we built a machine learning model based on the architecture of TransFuse [20] to accomplish the segmentation task. The TransFuse model combines Transformers and convolutional neural networks (CNNs), taking advantage of both local and global features. In this paper, we propose a modification to the data flow in encoder necks for decoding features in the higher resolution level, and in fusion modules for more efficient attention. Furthermore, to minimize the information loss as a result of resizing, we propose new techniques in both training and testing algorithms. Firstly, a region proposal network (RPN) is introduced from object detection methods and is used at the image pre-processing phase. It crops fixed size images from origin images, so that the high resolution input can be fed into TransFuse. We also applied test-time augmentation following a similar concept to RPN. We crop fixed size images at each corner and use edge pooling to ensemble them properly.

Original language	English
Title of host publication	Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings
Editors	Moi Hoon Yap, Connah Kendrick, Bill Cassidy
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	52-64
Number of pages	13
ISBN (Print)	9783031263538
DOIs	https://doi.org/10.1007/978-3-031-26354-5_5
State	Published - 2023
Event	3rd Diabetic Foot Ulcers Grand Challenge, DFUC 2022, held in Conjunction with the 25th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2022 - Singapore, Singapore Duration: 22 Sep 2022 → 22 Sep 2022

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	13797 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	3rd Diabetic Foot Ulcers Grand Challenge, DFUC 2022, held in Conjunction with the 25th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2022
Country/Territory	Singapore
City	Singapore
Period	22/09/22 → 22/09/22

Keywords

CNN
Ensemble
Medical image segmentation
Object detection
Transformer

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

Access to Document

10.1007/978-3-031-26354-5_5

Cite this

Chen, Y. H., Ju, Y. J., & Huang, J. D. (2023). Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images. In M. H. Yap, C. Kendrick, & B. Cassidy (Eds.), Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings (pp. 52-64). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 13797 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-26354-5_5

Chen, Yung Han ; Ju, Yi Jen ; Huang, Juinn Dar. / Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images. Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings. editor / Moi Hoon Yap ; Connah Kendrick ; Bill Cassidy. Springer Science and Business Media Deutschland GmbH, 2023. pp. 52-64 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{24a22d7dcede41f199244631aa37e133,

title = "Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images",

abstract = "For the Diabetic Foot Ulcer Challenge 2022 (DFUC2022) hosted by MICCAI 2022, we built a machine learning model based on the architecture of TransFuse [20] to accomplish the segmentation task. The TransFuse model combines Transformers and convolutional neural networks (CNNs), taking advantage of both local and global features. In this paper, we propose a modification to the data flow in encoder necks for decoding features in the higher resolution level, and in fusion modules for more efficient attention. Furthermore, to minimize the information loss as a result of resizing, we propose new techniques in both training and testing algorithms. Firstly, a region proposal network (RPN) is introduced from object detection methods and is used at the image pre-processing phase. It crops fixed size images from origin images, so that the high resolution input can be fed into TransFuse. We also applied test-time augmentation following a similar concept to RPN. We crop fixed size images at each corner and use edge pooling to ensemble them properly.",

keywords = "CNN, Ensemble, Medical image segmentation, Object detection, Transformer",

author = "Chen, {Yung Han} and Ju, {Yi Jen} and Huang, {Juinn Dar}",

note = "Publisher Copyright: {\textcopyright} 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.; 3rd Diabetic Foot Ulcers Grand Challenge, DFUC 2022, held in Conjunction with the 25th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2022 ; Conference date: 22-09-2022 Through 22-09-2022",

year = "2023",

doi = "10.1007/978-3-031-26354-5_5",

language = "English",

isbn = "9783031263538",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "52--64",

editor = "Yap, {Moi Hoon} and Connah Kendrick and Bill Cassidy",

booktitle = "Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings",

address = "德國",

}

Chen, YH, Ju, YJ & Huang, JD 2023, Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images. in MH Yap, C Kendrick & B Cassidy (eds), Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 13797 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 52-64, 3rd Diabetic Foot Ulcers Grand Challenge, DFUC 2022, held in Conjunction with the 25th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2022, Singapore, Singapore, 22/09/22. https://doi.org/10.1007/978-3-031-26354-5_5

Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images. / Chen, Yung Han; Ju, Yi Jen; Huang, Juinn Dar.
Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings. ed. / Moi Hoon Yap; Connah Kendrick; Bill Cassidy. Springer Science and Business Media Deutschland GmbH, 2023. p. 52-64 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 13797 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images

AU - Chen, Yung Han

AU - Ju, Yi Jen

AU - Huang, Juinn Dar

PY - 2023

Y1 - 2023

N2 - For the Diabetic Foot Ulcer Challenge 2022 (DFUC2022) hosted by MICCAI 2022, we built a machine learning model based on the architecture of TransFuse [20] to accomplish the segmentation task. The TransFuse model combines Transformers and convolutional neural networks (CNNs), taking advantage of both local and global features. In this paper, we propose a modification to the data flow in encoder necks for decoding features in the higher resolution level, and in fusion modules for more efficient attention. Furthermore, to minimize the information loss as a result of resizing, we propose new techniques in both training and testing algorithms. Firstly, a region proposal network (RPN) is introduced from object detection methods and is used at the image pre-processing phase. It crops fixed size images from origin images, so that the high resolution input can be fed into TransFuse. We also applied test-time augmentation following a similar concept to RPN. We crop fixed size images at each corner and use edge pooling to ensemble them properly.

AB - For the Diabetic Foot Ulcer Challenge 2022 (DFUC2022) hosted by MICCAI 2022, we built a machine learning model based on the architecture of TransFuse [20] to accomplish the segmentation task. The TransFuse model combines Transformers and convolutional neural networks (CNNs), taking advantage of both local and global features. In this paper, we propose a modification to the data flow in encoder necks for decoding features in the higher resolution level, and in fusion modules for more efficient attention. Furthermore, to minimize the information loss as a result of resizing, we propose new techniques in both training and testing algorithms. Firstly, a region proposal network (RPN) is introduced from object detection methods and is used at the image pre-processing phase. It crops fixed size images from origin images, so that the high resolution input can be fed into TransFuse. We also applied test-time augmentation following a similar concept to RPN. We crop fixed size images at each corner and use edge pooling to ensemble them properly.

KW - CNN

KW - Ensemble

KW - Medical image segmentation

KW - Object detection

KW - Transformer

UR - http://www.scopus.com/inward/record.url?scp=85151058807&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-26354-5_5

DO - 10.1007/978-3-031-26354-5_5

M3 - Conference contribution

AN - SCOPUS:85151058807

SN - 9783031263538

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 52

EP - 64

BT - Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings

A2 - Yap, Moi Hoon

A2 - Kendrick, Connah

A2 - Cassidy, Bill

PB - Springer Science and Business Media Deutschland GmbH

T2 - 3rd Diabetic Foot Ulcers Grand Challenge, DFUC 2022, held in Conjunction with the 25th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2022

Y2 - 22 September 2022 through 22 September 2022

ER -

Chen YH, Ju YJ, Huang JD. Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images. In Yap MH, Kendrick C, Cassidy B, editors, Diabetic Foot Ulcers Grand Challenge - 3rd Challenge, DFUC 2022, Held in Conjunction with MICCAI 2022, Proceedings. Springer Science and Business Media Deutschland GmbH. 2023. p. 52-64. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-031-26354-5_5