Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification

Rebecca Mes; Mark Wijkhuizen; Lennard M. van Karnenbeek; Tao Tan; Ritse Mann; Theo Ruers; Freija Geldof; Behdad Dashtbozorg

doi:10.1007/978-3-032-05559-0_3

Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification

Rebecca Mes
, Mark Wijkhuizen
, Lennard M. van Karnenbeek
, Tao Tan
, Ritse Mann
, Theo Ruers
, Freija Geldof
, Behdad Dashtbozorg^*

^*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

9 Downloads (Pure)

Abstract

Breast ultrasound (BUS) offers a low-cost, radiation-free imaging alternative for breast cancer diagnostics, particularly suitable for point-of-care use. Despite promising results from deep learning (DL) models for BUS lesion classification, most models show significant performance drops on external datasets, suggesting overfitting to dataset-specific features. This lack of generalizability is concerning, especially given disparities in breast cancer outcomes across demographic groups and the diversity of ultrasound acquisition conditions. In this study, we investigate the impact of training dataset diversity on the robustness of DL models for BUS lesion classification. We compare three model architectures: ResNet50 (CNN), MViTv2 (Vision Transformer), and MambaOut (Vision Mamba), using 8403 B-mode BUS images from ten publicly available datasets originating from seven countries. Models were evaluated under three scenarios: single-dataset training, leave-one-dataset-out (LODO), and limited-data all-source training. Our results show that performance strongly depends on training set composition, with certain datasets consistently yielding better model performance, and substantial variability in cross-dataset generalization. This study provides new insights into the design of fair and generalizable DL systems for breast cancer diagnostics.

Original language	English
Title of host publication	Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care
Subtitle of host publication	2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings
Editors	Tianyu Zhang, Oliver Lester Saldanha, Luyi Han, Nika Rasoolzadeh, Lidia Garrucho Moras, Jarek van Dijk, Tao Tan, Jakob Nikolas Kather, Ritse Mann
Publisher	Springer
Pages	21-30
ISBN (Electronic)	978-3-032-05559-0
ISBN (Print)	9783032055583
DOIs	https://doi.org/10.1007/978-3-032-05559-0_3
Publication status	Published - 2026
Externally published	Yes
Event	2nd Deep Breast Workshop on Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, Deep-Breath 2025 - Daejeon, Korea, Republic of Duration: 23 Sept 2025 → 23 Sept 2025 Conference number: 2 https://deep-breath-miccai.github.io/deepbreath-2025/

Publication series

Name	Lecture Notes in Computer Science
Volume	16142 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Workshop

Workshop	2nd Deep Breast Workshop on Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, Deep-Breath 2025
Abbreviated title	Deep-Breath 2025
Country/Territory	Korea, Republic of
City	Daejeon
Period	23/09/25 → 23/09/25
Internet address	https://deep-breath-miccai.github.io/deepbreath-2025/

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

SDG 3 Good Health and Well-being

Keywords

2026 OA procedure
Deep Learning (DL)
Domain generalization
Lesion classification
Ultrasound
Breast cancer
Data diversity

Access to Document

10.1007/978-3-032-05559-0_3

ArticleFinal published version, 1.31 MBLicence: Taverne

Cite this

Mes, R., Wijkhuizen, M., van Karnenbeek, L. M., Tan, T., Mann, R., Ruers, T., Geldof, F., & Dashtbozorg, B. (2026). Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification. In T. Zhang, O. L. Saldanha, L. Han, N. Rasoolzadeh, L. Garrucho Moras, J. van Dijk, T. Tan, J. N. Kather, & R. Mann (Eds.), Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care: 2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings (pp. 21-30). (Lecture Notes in Computer Science; Vol. 16142 LNCS). Springer. https://doi.org/10.1007/978-3-032-05559-0_3

Mes, Rebecca ; Wijkhuizen, Mark ; van Karnenbeek, Lennard M. et al. / Training Set Diversity : A Key Factor in AI-Driven Breast Ultrasound Classification. Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care: 2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings. editor / Tianyu Zhang ; Oliver Lester Saldanha ; Luyi Han ; Nika Rasoolzadeh ; Lidia Garrucho Moras ; Jarek van Dijk ; Tao Tan ; Jakob Nikolas Kather ; Ritse Mann. Springer, 2026. pp. 21-30 (Lecture Notes in Computer Science).

@inproceedings{5a9d71e6f1b94d72974b4fdcf43b4969,

title = "Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification",

abstract = "Breast ultrasound (BUS) offers a low-cost, radiation-free imaging alternative for breast cancer diagnostics, particularly suitable for point-of-care use. Despite promising results from deep learning (DL) models for BUS lesion classification, most models show significant performance drops on external datasets, suggesting overfitting to dataset-specific features. This lack of generalizability is concerning, especially given disparities in breast cancer outcomes across demographic groups and the diversity of ultrasound acquisition conditions. In this study, we investigate the impact of training dataset diversity on the robustness of DL models for BUS lesion classification. We compare three model architectures: ResNet50 (CNN), MViTv2 (Vision Transformer), and MambaOut (Vision Mamba), using 8403 B-mode BUS images from ten publicly available datasets originating from seven countries. Models were evaluated under three scenarios: single-dataset training, leave-one-dataset-out (LODO), and limited-data all-source training. Our results show that performance strongly depends on training set composition, with certain datasets consistently yielding better model performance, and substantial variability in cross-dataset generalization. This study provides new insights into the design of fair and generalizable DL systems for breast cancer diagnostics.",

keywords = "2026 OA procedure, Deep Learning (DL), Domain generalization, Lesion classification, Ultrasound, Breast cancer, Data diversity",

author = "Rebecca Mes and Mark Wijkhuizen and \{van Karnenbeek\}, \{Lennard M.\} and Tao Tan and Ritse Mann and Theo Ruers and Freija Geldof and Behdad Dashtbozorg",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.; 2nd Deep Breast Workshop on Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, Deep-Breath 2025, Deep-Breath 2025 ; Conference date: 23-09-2025 Through 23-09-2025",

year = "2026",

doi = "10.1007/978-3-032-05559-0\_3",

language = "English",

isbn = "9783032055583",

series = "Lecture Notes in Computer Science",

publisher = "Springer",

pages = "21--30",

editor = "Tianyu Zhang and Saldanha, \{Oliver Lester\} and Luyi Han and Nika Rasoolzadeh and \{Garrucho Moras\}, Lidia and \{van Dijk\}, Jarek and Tao Tan and Kather, \{Jakob Nikolas\} and Ritse Mann",

booktitle = "Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care",

address = "Germany",

url = "https://deep-breath-miccai.github.io/deepbreath-2025/",

}

Mes, R, Wijkhuizen, M, van Karnenbeek, LM, Tan, T, Mann, R, Ruers, T, Geldof, F & Dashtbozorg, B 2026, Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification. in T Zhang, OL Saldanha, L Han, N Rasoolzadeh, L Garrucho Moras, J van Dijk, T Tan, JN Kather & R Mann (eds), Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care: 2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings. Lecture Notes in Computer Science, vol. 16142 LNCS, Springer, pp. 21-30, 2nd Deep Breast Workshop on Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, Deep-Breath 2025, Daejeon, Korea, Republic of, 23/09/25. https://doi.org/10.1007/978-3-032-05559-0_3

Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification. / Mes, Rebecca; Wijkhuizen, Mark; van Karnenbeek, Lennard M. et al.
Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care: 2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings. ed. / Tianyu Zhang; Oliver Lester Saldanha; Luyi Han; Nika Rasoolzadeh; Lidia Garrucho Moras; Jarek van Dijk; Tao Tan; Jakob Nikolas Kather; Ritse Mann. Springer, 2026. p. 21-30 (Lecture Notes in Computer Science; Vol. 16142 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Training Set Diversity

T2 - 2nd Deep Breast Workshop on Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, Deep-Breath 2025

AU - Mes, Rebecca

AU - Wijkhuizen, Mark

AU - van Karnenbeek, Lennard M.

AU - Tan, Tao

AU - Mann, Ritse

AU - Ruers, Theo

AU - Geldof, Freija

AU - Dashtbozorg, Behdad

N1 - Conference code: 2

PY - 2026

Y1 - 2026

N2 - Breast ultrasound (BUS) offers a low-cost, radiation-free imaging alternative for breast cancer diagnostics, particularly suitable for point-of-care use. Despite promising results from deep learning (DL) models for BUS lesion classification, most models show significant performance drops on external datasets, suggesting overfitting to dataset-specific features. This lack of generalizability is concerning, especially given disparities in breast cancer outcomes across demographic groups and the diversity of ultrasound acquisition conditions. In this study, we investigate the impact of training dataset diversity on the robustness of DL models for BUS lesion classification. We compare three model architectures: ResNet50 (CNN), MViTv2 (Vision Transformer), and MambaOut (Vision Mamba), using 8403 B-mode BUS images from ten publicly available datasets originating from seven countries. Models were evaluated under three scenarios: single-dataset training, leave-one-dataset-out (LODO), and limited-data all-source training. Our results show that performance strongly depends on training set composition, with certain datasets consistently yielding better model performance, and substantial variability in cross-dataset generalization. This study provides new insights into the design of fair and generalizable DL systems for breast cancer diagnostics.

AB - Breast ultrasound (BUS) offers a low-cost, radiation-free imaging alternative for breast cancer diagnostics, particularly suitable for point-of-care use. Despite promising results from deep learning (DL) models for BUS lesion classification, most models show significant performance drops on external datasets, suggesting overfitting to dataset-specific features. This lack of generalizability is concerning, especially given disparities in breast cancer outcomes across demographic groups and the diversity of ultrasound acquisition conditions. In this study, we investigate the impact of training dataset diversity on the robustness of DL models for BUS lesion classification. We compare three model architectures: ResNet50 (CNN), MViTv2 (Vision Transformer), and MambaOut (Vision Mamba), using 8403 B-mode BUS images from ten publicly available datasets originating from seven countries. Models were evaluated under three scenarios: single-dataset training, leave-one-dataset-out (LODO), and limited-data all-source training. Our results show that performance strongly depends on training set composition, with certain datasets consistently yielding better model performance, and substantial variability in cross-dataset generalization. This study provides new insights into the design of fair and generalizable DL systems for breast cancer diagnostics.

KW - 2026 OA procedure

KW - Deep Learning (DL)

KW - Domain generalization

KW - Lesion classification

KW - Ultrasound

KW - Breast cancer

KW - Data diversity

UR - https://www.scopus.com/pages/publications/105018579764

U2 - 10.1007/978-3-032-05559-0_3

DO - 10.1007/978-3-032-05559-0_3

M3 - Conference contribution

AN - SCOPUS:105018579764

SN - 9783032055583

T3 - Lecture Notes in Computer Science

SP - 21

EP - 30

BT - Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care

A2 - Zhang, Tianyu

A2 - Saldanha, Oliver Lester

A2 - Han, Luyi

A2 - Rasoolzadeh, Nika

A2 - Garrucho Moras, Lidia

A2 - van Dijk, Jarek

A2 - Tan, Tao

A2 - Kather, Jakob Nikolas

A2 - Mann, Ritse

PB - Springer

Y2 - 23 September 2025 through 23 September 2025

ER -

Mes R, Wijkhuizen M, van Karnenbeek LM, Tan T, Mann R, Ruers T et al. Training Set Diversity: A Key Factor in AI-Driven Breast Ultrasound Classification. In Zhang T, Saldanha OL, Han L, Rasoolzadeh N, Garrucho Moras L, van Dijk J, Tan T, Kather JN, Mann R, editors, Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care: 2nd Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025, Proceedings. Springer. 2026. p. 21-30. (Lecture Notes in Computer Science). Epub 2025 Sept 21. doi: 10.1007/978-3-032-05559-0_3