Towards Out-of-Distribution Detection for Breast Cancer Classification in Point-of-Care Ultrasound Imaging
(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)
- author
- Karlsson, Jennie
LU
; Wodrich, Marisa
LU
; Overgaard, Niels Christian
LU
; Sahlin, Freja
; Lång, Kristina
LU
; Heyden, Anders
LU
and Arvidsson, Ida LU
- organization
-
- Computer Vision and Machine Learning (research group)
- LU Profile Area: Natural and Artificial Cognition
- eSSENCE: The e-Science Collaboration
- LTH Profile Area: AI and Digitalization
- LTH Profile Area: Engineering Health
- Engineering Mathematics (M.Sc.Eng.)
- Mathematical Imaging Group (research group)
- Partial differential equations (research group)
- LUCC: Lund University Cancer Centre
- Radiology Diagnostics, Malmö (research group)
- LU Profile Area: Proactive Ageing
- ELLIIT: the Linköping-Lund initiative on IT and mobile communication
- publishing date
- 2025
- 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}}, }