Sammanfattning

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%.

Originalspråkengelska
Titel på värdpublikationPattern Recognition - 27th International Conference, ICPR 2024, Proceedings, Part XIII
RedaktörerApostolos Antonacopoulos, Subhasis Chaudhuri, Rama Chellappa, Cheng-Lin Liu, Saumik Bhattacharya, Umapada Pal
FörlagSpringer Science and Business Media B.V.
Sidor49-63
ISBN (elektroniskt)978-3-031-78201-5
ISBN (tryckt)978-3-031-78200-8
DOI
StatusPublished - 2025
Evenemang27th International Conference on Pattern Recognition, ICPR 2024 - Kolkata, Indien
Varaktighet: 2024 dec. 12024 dec. 5

Publikationsserier

NamnLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volym15313 LNCS
ISSN (tryckt)0302-9743
ISSN (elektroniskt)1611-3349

Konferens

Konferens27th International Conference on Pattern Recognition, ICPR 2024
Land/TerritoriumIndien
OrtKolkata
Period2024/12/012024/12/05

Bibliografisk information

Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

Ämnesklassifikation (UKÄ)

  • Cancer och onkologi
  • Radiologi och bildbehandling

Citera det här