Trustworthiness for Deep Learning Based Breast Cancer Detection Using Point-of-Care Ultrasound Imaging in Low-Resource Settings
(2025) 1st MICCAI Meets Africa Workshop, MImA 2024 and 1st MICCAI Student Board Workshop on Empowering Medical Information Computing and Research through Early-Career Expertise, EMERGE 2024, Held in Conjunction with MICCAI 2024 In Communications in Computer and Information Science 2240. p.42-51- Abstract
Poor survival for breast cancer in low- and middle-income countries is largely attributed to late-stage diagnosis and limited access to diagnostic tools. Therefore, we propose using point-of-care ultrasound (POCUS) as a low-cost diagnostic solution paired with a deep learning (DL) model for cancer detection. While using DL has shown great potential, it is crucial to ensure the trustworthiness of a model for diagnostic support in a setting with minimally trained healthcare workers, as a wrong prediction can lead to severe consequences. In the present study, we investigated different measures of trustworthiness from the field of uncertainty quantification, including deep ensembles, Bayesian neural networks and softmax score, for the... (More)
Poor survival for breast cancer in low- and middle-income countries is largely attributed to late-stage diagnosis and limited access to diagnostic tools. Therefore, we propose using point-of-care ultrasound (POCUS) as a low-cost diagnostic solution paired with a deep learning (DL) model for cancer detection. While using DL has shown great potential, it is crucial to ensure the trustworthiness of a model for diagnostic support in a setting with minimally trained healthcare workers, as a wrong prediction can lead to severe consequences. In the present study, we investigated different measures of trustworthiness from the field of uncertainty quantification, including deep ensembles, Bayesian neural networks and softmax score, for the application of breast cancer detection in POCUS imaging. The results show that all methods exhibit a correlation between uncertainty scores and correctness of prediction. The correlation was strongest when using an average ensemble and an entropy-based total predictive uncertainty. When excluding 30% of the test samples based on highest uncertainty scores, the area under the receiver operating characteristic curve (AUC) for cancer detection increases significantly from 95.6% to 98.9% (with 95% confidence intervals [93.3, 97.0] to [97.5, 99.9]), comparable to an expert radiologist’s performance.
(Less)
- author
- Wodrich, Marisa
LU
; Karlsson, Jennie
LU
; Lång, Kristina
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
- Mathematics (Faculty of Engineering)
- LUCC: Lund University Cancer Centre
- Radiology Diagnostics, Malmö (research group)
- Tumor microenvironment
- LU Profile Area: Proactive Ageing
- LTH Profile Area: AI and Digitalization
- publishing date
- 2025
- type
- Chapter in Book/Report/Conference proceeding
- publication status
- published
- subject
- keywords
- Breast cancer detection, low-resource settings, Point-of-care ultrasound, Trustworthy AI, Uncertainty quantification
- host publication
- Medical Information Computing : First MICCAI Meets Africa Workshop, MImA 2024, and First MICCAI Student Board Workshop on Empowering Medical Information Computing and Research through Early-Career Expertise, EMERGE 2024, Held in Conjunction with MICCAI 2024, Revised Selected Papers - First MICCAI Meets Africa Workshop, MImA 2024, and First MICCAI Student Board Workshop on Empowering Medical Information Computing and Research through Early-Career Expertise, EMERGE 2024, Held in Conjunction with MICCAI 2024, Revised Selected Papers
- series title
- Communications in Computer and Information Science
- editor
- Anazodo, Udunna ; Akash, Naren ; Fuchs, Moritz ; Cintas, Celia ; Crimi, Alessandro ; Mutsvangwa, Tinahse ; Dako, Farouk and Ogallo, Willam
- volume
- 2240
- pages
- 42 - 51
- publisher
- Springer Science and Business Media B.V.
- conference name
- 1st MICCAI Meets Africa Workshop, MImA 2024 and 1st MICCAI Student Board Workshop on Empowering Medical Information Computing and Research through Early-Career Expertise, EMERGE 2024, Held in Conjunction with MICCAI 2024
- conference location
- Marrakesh, Morocco
- conference dates
- 2024-10-06 - 2024-10-06
- external identifiers
-
- scopus:85219195184
- ISSN
- 1865-0937
- 1865-0929
- ISBN
- 9783031791024
- DOI
- 10.1007/978-3-031-79103-1_5
- language
- English
- LU publication?
- yes
- additional info
- Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
- id
- 7faea94a-ade9-4f73-a32e-d5c3e8366f05
- date added to LUP
- 2025-03-15 08:07:50
- date last changed
- 2025-07-05 18:27:06
@inproceedings{7faea94a-ade9-4f73-a32e-d5c3e8366f05, abstract = {{<p>Poor survival for breast cancer in low- and middle-income countries is largely attributed to late-stage diagnosis and limited access to diagnostic tools. Therefore, we propose using point-of-care ultrasound (POCUS) as a low-cost diagnostic solution paired with a deep learning (DL) model for cancer detection. While using DL has shown great potential, it is crucial to ensure the trustworthiness of a model for diagnostic support in a setting with minimally trained healthcare workers, as a wrong prediction can lead to severe consequences. In the present study, we investigated different measures of trustworthiness from the field of uncertainty quantification, including deep ensembles, Bayesian neural networks and softmax score, for the application of breast cancer detection in POCUS imaging. The results show that all methods exhibit a correlation between uncertainty scores and correctness of prediction. The correlation was strongest when using an average ensemble and an entropy-based total predictive uncertainty. When excluding 30% of the test samples based on highest uncertainty scores, the area under the receiver operating characteristic curve (AUC) for cancer detection increases significantly from 95.6% to 98.9% (with 95% confidence intervals [93.3, 97.0] to [97.5, 99.9]), comparable to an expert radiologist’s performance.</p>}}, author = {{Wodrich, Marisa and Karlsson, Jennie and Lång, Kristina and Arvidsson, Ida}}, booktitle = {{Medical Information Computing : First MICCAI Meets Africa Workshop, MImA 2024, and First MICCAI Student Board Workshop on Empowering Medical Information Computing and Research through Early-Career Expertise, EMERGE 2024, Held in Conjunction with MICCAI 2024, Revised Selected Papers}}, editor = {{Anazodo, Udunna and Akash, Naren and Fuchs, Moritz and Cintas, Celia and Crimi, Alessandro and Mutsvangwa, Tinahse and Dako, Farouk and Ogallo, Willam}}, isbn = {{9783031791024}}, issn = {{1865-0937}}, keywords = {{Breast cancer detection; low-resource settings; Point-of-care ultrasound; Trustworthy AI; Uncertainty quantification}}, language = {{eng}}, pages = {{42--51}}, publisher = {{Springer Science and Business Media B.V.}}, series = {{Communications in Computer and Information Science}}, title = {{Trustworthiness for Deep Learning Based Breast Cancer Detection Using Point-of-Care Ultrasound Imaging in Low-Resource Settings}}, url = {{http://dx.doi.org/10.1007/978-3-031-79103-1_5}}, doi = {{10.1007/978-3-031-79103-1_5}}, volume = {{2240}}, year = {{2025}}, }