Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Towards Out-of-Distribution Detection for Breast Cancer Classification in Point-of-Care Ultrasound Imaging

Karlsson, Jennie LU ; Wodrich, Marisa LU ; Overgaard, Niels Christian LU ; Sahlin, Freja ; Lång, Kristina LU ; Heyden, Anders LU orcid and Arvidsson, Ida LU orcid (2025) 27th International Conference on Pattern Recognition, ICPR 2024 In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 15313 LNCS. p.49-63
Abstract

The use of deep learning for classification tasks has shown great potential in medical applications. In critical domains as such, it is of high interest to have trustworthy algorithms which are able to tell when a reliable assessment cannot be guaranteed. Hence, detecting out-of-distribution (OOD) samples is a crucial step towards building a safe classifier. Following a previous study, showing that it is possible to classify breast cancer in point-of-care ultrasound (POCUS) images, this study investigates out-of-distribution (OOD) detection. Three different OOD detection methods were implemented and evaluated in this study: softmax score, multi-level energy score and deep ensembles. As in-distribution training data both standard... (More)

The use of deep learning for classification tasks has shown great potential in medical applications. In critical domains as such, it is of high interest to have trustworthy algorithms which are able to tell when a reliable assessment cannot be guaranteed. Hence, detecting out-of-distribution (OOD) samples is a crucial step towards building a safe classifier. Following a previous study, showing that it is possible to classify breast cancer in point-of-care ultrasound (POCUS) images, this study investigates out-of-distribution (OOD) detection. Three different OOD detection methods were implemented and evaluated in this study: softmax score, multi-level energy score and deep ensembles. As in-distribution training data both standard ultrasound images and POCUS images were used and a separate POCUS data set was used for testing. All OOD detection methods were evaluated on three different OOD data sets, which are a mixture of synthetic data and real ultrasound data that represent different use cases for which OOD detection in automatic breast cancer classification is needed, covering a range of simple OOD cases, ultrasound images of poor quality and ultrasound images of non-breast tissue. The results show that the softmax score is inferior compared to the other methods at detecting OOD samples. The multi-level energy score performs superior on two of the OOD data sets. The deep ensembles perform superior on the OOD data set containing ultrasound images of poor quality with a 95% confidence interval for the area under the receiver operating characteristic curve of 97.2%–98.5%.

(Less)
Please use this url to cite or link to this publication:
author
; ; ; ; ; and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
breast cancer classification, Out-of-distribution detection, point-of-care ultrasound
host publication
Pattern Recognition - 27th International Conference, ICPR 2024, Proceedings, Part XIII
series title
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
editor
Antonacopoulos, Apostolos ; Chaudhuri, Subhasis ; Chellappa, Rama ; Liu, Cheng-Lin ; Bhattacharya, Saumik and Pal, Umapada
volume
15313 LNCS
pages
49 - 63
publisher
Springer Science and Business Media B.V.
conference name
27th International Conference on Pattern Recognition, ICPR 2024
conference location
Kolkata, India
conference dates
2024-12-01 - 2024-12-05
external identifiers
  • scopus:85211806616
ISSN
1611-3349
0302-9743
ISBN
978-3-031-78201-5
978-3-031-78200-8
DOI
10.1007/978-3-031-78201-5_4
language
English
LU publication?
yes
additional info
Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
id
1add9dcf-8b02-4756-988d-126a4cdd5ff4
date added to LUP
2024-12-21 08:13:06
date last changed
2025-07-06 00:17:57
@inproceedings{1add9dcf-8b02-4756-988d-126a4cdd5ff4,
  abstract     = {{<p>The use of deep learning for classification tasks has shown great potential in medical applications. In critical domains as such, it is of high interest to have trustworthy algorithms which are able to tell when a reliable assessment cannot be guaranteed. Hence, detecting out-of-distribution (OOD) samples is a crucial step towards building a safe classifier. Following a previous study, showing that it is possible to classify breast cancer in point-of-care ultrasound (POCUS) images, this study investigates out-of-distribution (OOD) detection. Three different OOD detection methods were implemented and evaluated in this study: softmax score, multi-level energy score and deep ensembles. As in-distribution training data both standard ultrasound images and POCUS images were used and a separate POCUS data set was used for testing. All OOD detection methods were evaluated on three different OOD data sets, which are a mixture of synthetic data and real ultrasound data that represent different use cases for which OOD detection in automatic breast cancer classification is needed, covering a range of simple OOD cases, ultrasound images of poor quality and ultrasound images of non-breast tissue. The results show that the softmax score is inferior compared to the other methods at detecting OOD samples. The multi-level energy score performs superior on two of the OOD data sets. The deep ensembles perform superior on the OOD data set containing ultrasound images of poor quality with a 95% confidence interval for the area under the receiver operating characteristic curve of 97.2%–98.5%.</p>}},
  author       = {{Karlsson, Jennie and Wodrich, Marisa and Overgaard, Niels Christian and Sahlin, Freja and Lång, Kristina and Heyden, Anders and Arvidsson, Ida}},
  booktitle    = {{Pattern Recognition - 27th International Conference, ICPR 2024, Proceedings, Part XIII}},
  editor       = {{Antonacopoulos, Apostolos and Chaudhuri, Subhasis and Chellappa, Rama and Liu, Cheng-Lin and Bhattacharya, Saumik and Pal, Umapada}},
  isbn         = {{978-3-031-78201-5}},
  issn         = {{1611-3349}},
  keywords     = {{breast cancer classification; Out-of-distribution detection; point-of-care ultrasound}},
  language     = {{eng}},
  pages        = {{49--63}},
  publisher    = {{Springer Science and Business Media B.V.}},
  series       = {{Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)}},
  title        = {{Towards Out-of-Distribution Detection for Breast Cancer Classification in Point-of-Care Ultrasound Imaging}},
  url          = {{http://dx.doi.org/10.1007/978-3-031-78201-5_4}},
  doi          = {{10.1007/978-3-031-78201-5_4}},
  volume       = {{15313 LNCS}},
  year         = {{2025}},
}