In Silico Digital Breast Tomosynthesis Dataset for the Comparative Analysis of Deep Learning Models in Tumor Segmentation

Cristina Alfaro Vergara; Nicolás Araya Caro; Domingo Mery Quiroz; Claudia Prieto Vasquez

doi:10.1007/s10278-025-01626-z

In Silico Digital Breast Tomosynthesis Dataset for the Comparative Analysis of Deep Learning Models in Tumor Segmentation

Título traducido de la contribución: Conjunto de datos de tomosíntesis mamaria digital in silico para el análisis comparativo de modelos de aprendizaje profundo en la segmentación tumoral

Cristina Alfaro Vergara
, Nicolás Araya Caro
, Domingo Mery Quiroz
, Claudia Prieto Vasquez

Departamento de Tecnología Médica

Pontificia Universidad Católica de Chile

Producción científica: Contribución a una revista › Artículo › revisión exhaustiva

Resumen

La escasez de conjuntos de datos de tomosíntesis digital mamaria (DBT) disponibles públicamente limita significativamente el desarrollo de modelos robustos de aprendizaje profundo (DL) para la segmentación de tumores mamarios. En este estudio exploratorio de prueba de concepto, evaluamos la viabilidad de los datos de DBT generados in silico como fuente de entrenamiento para la segmentación tumoral. Se utilizó un conjunto de datos de 230 regiones de interés (ROI) bidimensionales (2D) derivadas de software aprobado por la FDA y que abarcan un espectro de densidades mamarias y complejidades tumorales, para entrenar 13 modelos de DL, incluyendo arquitecturas U-Net, FCN, DeepLabv3 y DeepLabv3+. Cada modelo se entrenó desde cero o se ajustó utilizando pesos preentrenados por COCO (columnas vertebrales ResNet50/101). El rendimiento se evaluó mediante la puntuación F1, la intersección sobre la unión (IoU), la precisión y la recuperación. Entre todos los modelos, U-Net entrenado desde cero y DeepLabv3 + ajustado con ResNet50 lograron los resultados más altos y consistentes (puntuaciones F1 de 82,52% y 84,98%, e IoU por imagen de 78,49% y 83,77%, respectivamente). No se encontraron diferencias estadísticamente significativas utilizando la prueba de rangos con signo de Wilcoxon y la corrección de Bonferroni post hoc (α > 0,0042). Para evaluar la generalización entre dominios, el modelo U-Net de referencia se volvió a entrenar desde cero en un conjunto de datos híbrido que combinaba ROI DBT in silico y del mundo real, lo que produjo resultados prometedores (puntuación F1 de 79%). A pesar del cambio de dominio, estos hallazgos respaldan la utilidad de la DBT in silico como un recurso complementario para entrenar y comparar modelos DL, particularmente en entornos con datos limitados. Este estudio proporciona evidencia experimental fundamental para integrar datos in silico generados computacionalmente en flujos de trabajo de investigación de segmentación tumoral DBT basados en IA.

Título traducido de la contribución	Conjunto de datos de tomosíntesis mamaria digital in silico para el análisis comparativo de modelos de aprendizaje profundo en la segmentación tumoral
Idioma original	Inglés
Páginas (desde-hasta)	1-17
Número de páginas	17
Publicación	Journal of Imaging Informatics in Medicine
DOI	https://doi.org/10.1007/s10278-025-01626-z
Estado	Publicada - 4 ago. 2025

Palabras clave

Segmentación de tumores mamarios
Aprendizaje profundo U-Net
Tomosíntesis mamaria digital (DBT)
Entrenamiento híbrido
Datos in silico

Acceder al documento

10.1007/s10278-025-01626-z

Otros archivos y enlaces

Enlace a la publicación en Scopus

Citar esto

@article{f7465de65b694fd4b8872ce536354df5,

title = "In Silico Digital Breast Tomosynthesis Dataset for the Comparative Analysis of Deep Learning Models in Tumor Segmentation",

abstract = "The scarcity of publicly available digital breast tomosynthesis (DBT) datasets significantly limits the development of robust deep learning (DL) models for breast tumor segmentation. In this exploratory proof-of-concept study, we assess the viability of in silico-generated DBT data as a training source for tumor segmentation. A dataset of 230 two-dimensional (2D) regions of interest (ROIs) derived from FDA-cleared software and encompassing a spectrum of breast densities and tumor complexities, was used to train 13 DL models, including U-Net, FCN, DeepLabv3, and DeepLabv3 + architectures. Each model was trained either from scratch or fine-tuned using COCO-pretrained weights (ResNet50/101 backbones). Performance was evaluated using F1-score, intersection over union (IoU), precision, and recall. Among all models, U-Net trained from scratch and DeepLabv3 + fine-tuned with ResNet50 achieved the highest and most consistent results (F1-scores of 82.52\% and 84.98\%, and per-image IoUs of 78.49\% and 83.77\%, respectively). No statistically significant differences were found using the Wilcoxon signed-rank test and post hoc Bonferroni correction (α > 0.0042). To evaluate generalization across domains, the baseline U-Net model was retrained from scratch on a hybrid dataset combining in silico and real-world DBT ROIs, yielding promising results (F1-score of 79\%). Despite the domain shift, these findings support the utility of in silico DBT as a complementary resource for training and benchmarking DL models, particularly in data-limited environments. This study provides foundational experimental evidence for integrating computationally generated in silico data into AI-based DBT tumor segmentation research workflows.",

keywords = "Breast tumor segmentation, Deep learning U-Net, Digital breast tomosynthesis (DBT), Hybrid training, In silico data, Segmentaci{\'o}n de tumores mamarios, Aprendizaje profundo U-Net, Tomos{\'i}ntesis mamaria digital (DBT), Entrenamiento h{\'i}brido, Datos in silico",

author = "\{Alfaro Vergara\}, Cristina and \{Araya Caro\}, Nicol{\'a}s and \{Mery Quiroz\}, Domingo and \{Prieto Vasquez\}, Claudia",

note = "Publisher Copyright: {\textcopyright} The Author(s) 2025.",

year = "2025",

month = aug,

day = "4",

doi = "10.1007/s10278-025-01626-z",

language = "English",

pages = "1--17",

journal = "Journal of Imaging Informatics in Medicine",

issn = "2948-2933",

publisher = "Springer Nature",

}

TY - JOUR

T1 - In Silico Digital Breast Tomosynthesis Dataset for the Comparative Analysis of Deep Learning Models in Tumor Segmentation

AU - Alfaro Vergara, Cristina

AU - Araya Caro, Nicolás

AU - Mery Quiroz, Domingo

AU - Prieto Vasquez, Claudia

N1 - Publisher Copyright: © The Author(s) 2025.

PY - 2025/8/4

Y1 - 2025/8/4

N2 - The scarcity of publicly available digital breast tomosynthesis (DBT) datasets significantly limits the development of robust deep learning (DL) models for breast tumor segmentation. In this exploratory proof-of-concept study, we assess the viability of in silico-generated DBT data as a training source for tumor segmentation. A dataset of 230 two-dimensional (2D) regions of interest (ROIs) derived from FDA-cleared software and encompassing a spectrum of breast densities and tumor complexities, was used to train 13 DL models, including U-Net, FCN, DeepLabv3, and DeepLabv3 + architectures. Each model was trained either from scratch or fine-tuned using COCO-pretrained weights (ResNet50/101 backbones). Performance was evaluated using F1-score, intersection over union (IoU), precision, and recall. Among all models, U-Net trained from scratch and DeepLabv3 + fine-tuned with ResNet50 achieved the highest and most consistent results (F1-scores of 82.52% and 84.98%, and per-image IoUs of 78.49% and 83.77%, respectively). No statistically significant differences were found using the Wilcoxon signed-rank test and post hoc Bonferroni correction (α > 0.0042). To evaluate generalization across domains, the baseline U-Net model was retrained from scratch on a hybrid dataset combining in silico and real-world DBT ROIs, yielding promising results (F1-score of 79%). Despite the domain shift, these findings support the utility of in silico DBT as a complementary resource for training and benchmarking DL models, particularly in data-limited environments. This study provides foundational experimental evidence for integrating computationally generated in silico data into AI-based DBT tumor segmentation research workflows.

AB - The scarcity of publicly available digital breast tomosynthesis (DBT) datasets significantly limits the development of robust deep learning (DL) models for breast tumor segmentation. In this exploratory proof-of-concept study, we assess the viability of in silico-generated DBT data as a training source for tumor segmentation. A dataset of 230 two-dimensional (2D) regions of interest (ROIs) derived from FDA-cleared software and encompassing a spectrum of breast densities and tumor complexities, was used to train 13 DL models, including U-Net, FCN, DeepLabv3, and DeepLabv3 + architectures. Each model was trained either from scratch or fine-tuned using COCO-pretrained weights (ResNet50/101 backbones). Performance was evaluated using F1-score, intersection over union (IoU), precision, and recall. Among all models, U-Net trained from scratch and DeepLabv3 + fine-tuned with ResNet50 achieved the highest and most consistent results (F1-scores of 82.52% and 84.98%, and per-image IoUs of 78.49% and 83.77%, respectively). No statistically significant differences were found using the Wilcoxon signed-rank test and post hoc Bonferroni correction (α > 0.0042). To evaluate generalization across domains, the baseline U-Net model was retrained from scratch on a hybrid dataset combining in silico and real-world DBT ROIs, yielding promising results (F1-score of 79%). Despite the domain shift, these findings support the utility of in silico DBT as a complementary resource for training and benchmarking DL models, particularly in data-limited environments. This study provides foundational experimental evidence for integrating computationally generated in silico data into AI-based DBT tumor segmentation research workflows.

KW - Breast tumor segmentation

KW - Deep learning U-Net

KW - Digital breast tomosynthesis (DBT)

KW - Hybrid training

KW - In silico data

KW - Segmentación de tumores mamarios

KW - Aprendizaje profundo U-Net

KW - Tomosíntesis mamaria digital (DBT)

KW - Entrenamiento híbrido

KW - Datos in silico

UR - https://www.scopus.com/pages/publications/105012593610

U2 - 10.1007/s10278-025-01626-z

DO - 10.1007/s10278-025-01626-z

M3 - Article

AN - SCOPUS:105012593610

SN - 2948-2933

SP - 1

EP - 17

JO - Journal of Imaging Informatics in Medicine

JF - Journal of Imaging Informatics in Medicine

ER -

In Silico Digital Breast Tomosynthesis Dataset for the Comparative Analysis of Deep Learning Models in Tumor Segmentation

Resumen

Palabras clave

Acceder al documento

Otros archivos y enlaces

Huella

Citar esto