Tree-loss function for training neural networks on weakly-labelled datasets

Sergey Demyanov, Rajib Chakravorty, Zongyuan Ge, Seyedbehzad Bozorgtabar, Michelle Pablo, Adrian Bowling, Rahil Garnavi

Research output: Chapter in Book/Report/Conference proceedingConference PaperResearchpeer-review

1 Citation (Scopus)

Abstract

Neural networks are powerful tools for medical image classification and segmentation. However, existing network structures and training procedures assume that the output classes are mutually exclusive and equally important. Many datasets of medical images do not satisfy these conditions. For example, some skin disease datasets have images labelled as coarse-grained class (such as Benign) in addition to images with fine-grained labels (such as a Benign subclass called Blue Nevus), and conventional neural network can not leverage such additional data for training. Also, in the clinical decision making, some classes (such as skin cancer or Melanoma) often carry more importance than other lesion types. We propose a novel Tree-Loss function for training and fine-tuning a neural network classifier using all available labelled images. The key step is the definition of the class taxonomy tree, which is used to describe the relations between labels. The tree can be also adjusted to reflect the desired importance of each class. These steps can be performed by a domain expert without detailed knowledge of machine learning techniques. The experiments demonstrate the improved performance compared with the conventional approach even without using additional data.

Original languageEnglish
Title of host publication2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017)
EditorsSimon Warfield, Arrete Munoz-Barrutia
Place of PublicationPiscataway NJ USA
PublisherIEEE, Institute of Electrical and Electronics Engineers
Pages287-291
Number of pages5
ISBN (Electronic)9781509011728
ISBN (Print)9781509011735
DOIs
Publication statusPublished - 15 Jun 2017
Externally publishedYes
EventIEEE International Symposium on Biomedical Imaging (ISBI) 2017 - Melbourne Convention and Exhibition Centre, Melbourne, Australia
Duration: 18 Apr 201721 Apr 2017
Conference number: 14th
http://biomedicalimaging.org/2017/

Conference

ConferenceIEEE International Symposium on Biomedical Imaging (ISBI) 2017
Abbreviated titleISBI 2017
CountryAustralia
CityMelbourne
Period18/04/1721/04/17
OtherISBI is a joint initiative from the IEEE Signal Processing Society (SPS) and the IEEE Engineering in Medicine and Biology Society (EMBS). The 2017 meeting will include tutorials, and a scientific program composed of plenary talks, invited special sessions, challenges, as well as oral and poster presentations of peer-reviewed papers.
Internet address

Keywords

  • Classification
  • Loss function
  • Neural networks
  • Skin disease recognition
  • Taxonomy
  • Tree

Cite this

Demyanov, S., Chakravorty, R., Ge, Z., Bozorgtabar, S., Pablo, M., Bowling, A., & Garnavi, R. (2017). Tree-loss function for training neural networks on weakly-labelled datasets. In S. Warfield, & A. Munoz-Barrutia (Eds.), 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017) (pp. 287-291). [7950521] Piscataway NJ USA: IEEE, Institute of Electrical and Electronics Engineers. https://doi.org/10.1109/ISBI.2017.7950521
Demyanov, Sergey ; Chakravorty, Rajib ; Ge, Zongyuan ; Bozorgtabar, Seyedbehzad ; Pablo, Michelle ; Bowling, Adrian ; Garnavi, Rahil. / Tree-loss function for training neural networks on weakly-labelled datasets. 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). editor / Simon Warfield ; Arrete Munoz-Barrutia. Piscataway NJ USA : IEEE, Institute of Electrical and Electronics Engineers, 2017. pp. 287-291
@inproceedings{84ade4bcb3324439ab5760bb4b080495,
title = "Tree-loss function for training neural networks on weakly-labelled datasets",
abstract = "Neural networks are powerful tools for medical image classification and segmentation. However, existing network structures and training procedures assume that the output classes are mutually exclusive and equally important. Many datasets of medical images do not satisfy these conditions. For example, some skin disease datasets have images labelled as coarse-grained class (such as Benign) in addition to images with fine-grained labels (such as a Benign subclass called Blue Nevus), and conventional neural network can not leverage such additional data for training. Also, in the clinical decision making, some classes (such as skin cancer or Melanoma) often carry more importance than other lesion types. We propose a novel Tree-Loss function for training and fine-tuning a neural network classifier using all available labelled images. The key step is the definition of the class taxonomy tree, which is used to describe the relations between labels. The tree can be also adjusted to reflect the desired importance of each class. These steps can be performed by a domain expert without detailed knowledge of machine learning techniques. The experiments demonstrate the improved performance compared with the conventional approach even without using additional data.",
keywords = "Classification, Loss function, Neural networks, Skin disease recognition, Taxonomy, Tree",
author = "Sergey Demyanov and Rajib Chakravorty and Zongyuan Ge and Seyedbehzad Bozorgtabar and Michelle Pablo and Adrian Bowling and Rahil Garnavi",
year = "2017",
month = "6",
day = "15",
doi = "10.1109/ISBI.2017.7950521",
language = "English",
isbn = "9781509011735",
pages = "287--291",
editor = "Simon Warfield and Munoz-Barrutia, {Arrete }",
booktitle = "2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017)",
publisher = "IEEE, Institute of Electrical and Electronics Engineers",
address = "United States of America",

}

Demyanov, S, Chakravorty, R, Ge, Z, Bozorgtabar, S, Pablo, M, Bowling, A & Garnavi, R 2017, Tree-loss function for training neural networks on weakly-labelled datasets. in S Warfield & A Munoz-Barrutia (eds), 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017)., 7950521, IEEE, Institute of Electrical and Electronics Engineers, Piscataway NJ USA, pp. 287-291, IEEE International Symposium on Biomedical Imaging (ISBI) 2017, Melbourne, Australia, 18/04/17. https://doi.org/10.1109/ISBI.2017.7950521

Tree-loss function for training neural networks on weakly-labelled datasets. / Demyanov, Sergey; Chakravorty, Rajib; Ge, Zongyuan; Bozorgtabar, Seyedbehzad; Pablo, Michelle; Bowling, Adrian; Garnavi, Rahil.

2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). ed. / Simon Warfield; Arrete Munoz-Barrutia. Piscataway NJ USA : IEEE, Institute of Electrical and Electronics Engineers, 2017. p. 287-291 7950521.

Research output: Chapter in Book/Report/Conference proceedingConference PaperResearchpeer-review

TY - GEN

T1 - Tree-loss function for training neural networks on weakly-labelled datasets

AU - Demyanov, Sergey

AU - Chakravorty, Rajib

AU - Ge, Zongyuan

AU - Bozorgtabar, Seyedbehzad

AU - Pablo, Michelle

AU - Bowling, Adrian

AU - Garnavi, Rahil

PY - 2017/6/15

Y1 - 2017/6/15

N2 - Neural networks are powerful tools for medical image classification and segmentation. However, existing network structures and training procedures assume that the output classes are mutually exclusive and equally important. Many datasets of medical images do not satisfy these conditions. For example, some skin disease datasets have images labelled as coarse-grained class (such as Benign) in addition to images with fine-grained labels (such as a Benign subclass called Blue Nevus), and conventional neural network can not leverage such additional data for training. Also, in the clinical decision making, some classes (such as skin cancer or Melanoma) often carry more importance than other lesion types. We propose a novel Tree-Loss function for training and fine-tuning a neural network classifier using all available labelled images. The key step is the definition of the class taxonomy tree, which is used to describe the relations between labels. The tree can be also adjusted to reflect the desired importance of each class. These steps can be performed by a domain expert without detailed knowledge of machine learning techniques. The experiments demonstrate the improved performance compared with the conventional approach even without using additional data.

AB - Neural networks are powerful tools for medical image classification and segmentation. However, existing network structures and training procedures assume that the output classes are mutually exclusive and equally important. Many datasets of medical images do not satisfy these conditions. For example, some skin disease datasets have images labelled as coarse-grained class (such as Benign) in addition to images with fine-grained labels (such as a Benign subclass called Blue Nevus), and conventional neural network can not leverage such additional data for training. Also, in the clinical decision making, some classes (such as skin cancer or Melanoma) often carry more importance than other lesion types. We propose a novel Tree-Loss function for training and fine-tuning a neural network classifier using all available labelled images. The key step is the definition of the class taxonomy tree, which is used to describe the relations between labels. The tree can be also adjusted to reflect the desired importance of each class. These steps can be performed by a domain expert without detailed knowledge of machine learning techniques. The experiments demonstrate the improved performance compared with the conventional approach even without using additional data.

KW - Classification

KW - Loss function

KW - Neural networks

KW - Skin disease recognition

KW - Taxonomy

KW - Tree

UR - http://www.scopus.com/inward/record.url?scp=85023192515&partnerID=8YFLogxK

U2 - 10.1109/ISBI.2017.7950521

DO - 10.1109/ISBI.2017.7950521

M3 - Conference Paper

SN - 9781509011735

SP - 287

EP - 291

BT - 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017)

A2 - Warfield, Simon

A2 - Munoz-Barrutia, Arrete

PB - IEEE, Institute of Electrical and Electronics Engineers

CY - Piscataway NJ USA

ER -

Demyanov S, Chakravorty R, Ge Z, Bozorgtabar S, Pablo M, Bowling A et al. Tree-loss function for training neural networks on weakly-labelled datasets. In Warfield S, Munoz-Barrutia A, editors, 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017). Piscataway NJ USA: IEEE, Institute of Electrical and Electronics Engineers. 2017. p. 287-291. 7950521 https://doi.org/10.1109/ISBI.2017.7950521