Preoperative Nodal Metastatic Risk Evaluation in Early Breast Cancer : A Multimodal Deep Learning Approach
(2026)- Abstract
- Sentinel lymph node biopsy (SLNB) remains the standard method for axillary staging in breast cancer. However,
for the majority of patients with early breast cancer, it offers no treatment benefit and carries a risk of long-term
side effects. In the era of surgical de-escalation, reliable preoperative assessment of axillary lymph node metastasis is
essential to avoid overtreatment and support treatment planning in early-stage disease. The ever-increasing volume
of clinical data and biomedical resources, capturing multiple facets of the disease, presents new opportunities for
predicting nodal status. As a data-driven approach, deep learning has advanced significantly and shows strong
potential to improve predictive... (More) - Sentinel lymph node biopsy (SLNB) remains the standard method for axillary staging in breast cancer. However,
for the majority of patients with early breast cancer, it offers no treatment benefit and carries a risk of long-term
side effects. In the era of surgical de-escalation, reliable preoperative assessment of axillary lymph node metastasis is
essential to avoid overtreatment and support treatment planning in early-stage disease. The ever-increasing volume
of clinical data and biomedical resources, capturing multiple facets of the disease, presents new opportunities for
predicting nodal status. As a data-driven approach, deep learning has advanced significantly and shows strong
potential to improve predictive performance. This thesis aims to enhance preoperative assessment of axillary
lymph node status by integrating deep learning techniques with diverse clinical data.
The first study evaluates advanced deep learning algorithms for modeling tabular clinicopathological data to
predict nodal status. Surprisingly, none of the deep learning models show a significant advantage over a simple
linear regression baseline, suggesting the need for additional features. Thus, the second study incorporates gene
expression profiles with clinicopathological characteristics. The combined model improves prediction compared
to clinical data alone. Transformer architecture shows particular strength in extracting genomic representations
at scales and outperforms a prior biologically informed network. In the final study, a hybrid Transformer-ResNet
model is developed for high-resolution mammogram analysis, demonstrating that routine mammograms can pro-
vide substantial predictive information for nodal status. The results highlight the model’s ability to capture global
mammographic features and show that self-supervised transfer learning significantly enhances training perfor-
mance.
In conclusion, deep learning, particularly Transformer-based approaches, is well-suited for modeling diverse
biomedical data. Integrating clinical variables, mammographic images, and gene expression profiles offers a
promising strategy to improve preoperative assessment of axillary lymph nodes. This thesis provides a transla-
tional framework for developing and validating deep learning-based tools for nodal metastatic risk stratification to
guide clinical decision-making. It may support personalized care by identifying patients at ultra-low risk of nodal
burden to eliminate unnecessary SLNB, while preserving the ability to identify high-risk patients who would ben-
efit from surgical axillary staging. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/500f2e32-3b89-4679-a2c9-f4bee747d35e
- author
- Zhang, Daqu LU
- supervisor
-
- Patrik Edén LU
- Mattias Ohlsson LU
- Looket Dihge LU
- opponent
-
- Docent Eilertsen, Gabriel, Linköpings universitet.
- organization
- publishing date
- 2026
- type
- Thesis
- publication status
- published
- subject
- keywords
- deep learning, Prediction model
- pages
- 197 pages
- publisher
- Lund University, Faculty of Science
- defense location
- Världen, Sal GC1:F111, Geocentrum I
- defense date
- 2026-06-10 09:00:00
- ISBN
- 978-91-8104-975-6
- 978-91-8104-976-3
- project
- Applications of Deep Learning in Breast Cancer Research
- language
- English
- LU publication?
- yes
- id
- 500f2e32-3b89-4679-a2c9-f4bee747d35e
- date added to LUP
- 2026-05-05 12:06:57
- date last changed
- 2026-05-12 15:29:01
@phdthesis{500f2e32-3b89-4679-a2c9-f4bee747d35e,
abstract = {{Sentinel lymph node biopsy (SLNB) remains the standard method for axillary staging in breast cancer. However,<br/>for the majority of patients with early breast cancer, it offers no treatment benefit and carries a risk of long-term<br/>side effects. In the era of surgical de-escalation, reliable preoperative assessment of axillary lymph node metastasis is<br/>essential to avoid overtreatment and support treatment planning in early-stage disease. The ever-increasing volume<br/>of clinical data and biomedical resources, capturing multiple facets of the disease, presents new opportunities for<br/>predicting nodal status. As a data-driven approach, deep learning has advanced significantly and shows strong<br/>potential to improve predictive performance. This thesis aims to enhance preoperative assessment of axillary<br/>lymph node status by integrating deep learning techniques with diverse clinical data.<br/>The first study evaluates advanced deep learning algorithms for modeling tabular clinicopathological data to<br/>predict nodal status. Surprisingly, none of the deep learning models show a significant advantage over a simple<br/>linear regression baseline, suggesting the need for additional features. Thus, the second study incorporates gene<br/>expression profiles with clinicopathological characteristics. The combined model improves prediction compared<br/>to clinical data alone. Transformer architecture shows particular strength in extracting genomic representations<br/>at scales and outperforms a prior biologically informed network. In the final study, a hybrid Transformer-ResNet<br/>model is developed for high-resolution mammogram analysis, demonstrating that routine mammograms can pro-<br/>vide substantial predictive information for nodal status. The results highlight the model’s ability to capture global<br/>mammographic features and show that self-supervised transfer learning significantly enhances training perfor-<br/>mance.<br/>In conclusion, deep learning, particularly Transformer-based approaches, is well-suited for modeling diverse<br/>biomedical data. Integrating clinical variables, mammographic images, and gene expression profiles offers a<br/>promising strategy to improve preoperative assessment of axillary lymph nodes. This thesis provides a transla-<br/>tional framework for developing and validating deep learning-based tools for nodal metastatic risk stratification to<br/>guide clinical decision-making. It may support personalized care by identifying patients at ultra-low risk of nodal<br/>burden to eliminate unnecessary SLNB, while preserving the ability to identify high-risk patients who would ben-<br/>efit from surgical axillary staging.}},
author = {{Zhang, Daqu}},
isbn = {{978-91-8104-975-6}},
keywords = {{deep learning; Prediction model}},
language = {{eng}},
publisher = {{Lund University, Faculty of Science}},
school = {{Lund University}},
title = {{Preoperative Nodal Metastatic Risk Evaluation in Early Breast Cancer : A Multimodal Deep Learning Approach}},
url = {{https://lup.lub.lu.se/search/files/249270433/Daqu_Zhang_Doctoral_Thesis.pdf}},
year = {{2026}},
}