Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations

Lidia Garrucho, Kaisar Kushibar, Claire-Anne Reidel, Smriti Joshi, Richard Osuala, Apostolia Tsirikoglou, Maciej Bobowicz, Javier del Riego, Alessandro Catanese, Katarzyna Gwoździewicz, Maria-Laura Cosaka, Pasant M. Abo-Elhoda, Sara W. Tantawy, Shorouq S. Sakrana, Norhan O. Shawky-Abdelfatah, Amr Muhammad Abdo-Salem, Androniki Kozana, Eugen Divjak, Gordana Ivanac, Katerina Nikiforaki, Michail E. Klontzas, Rosa García-Dosdá, Meltem Gulsun-Akpinar, Oğuz Lafcı, Ritse Mann, Carlos Martín-Isla, Fred Prior, Kostas Marias, Martijn P. A. Starmans, Fredrik Strand, Oliver Díaz, Laura Igual, Karim Lekadir

TL;DR

This work delivers MAMA-MIA, the largest public multicenter breast cancer DCE-MRI benchmark with expert segmentations (1506 pre-treatment cases) by harmonizing four TCIA collections and generating 1506 expert masks through correction of preliminary nnU-Net segmentations. It provides 3D tumor and non-m mass-enhanced region annotations, 49 harmonized clinical/demographic variables, and pretrained baseline nnU-Net weights, enabling robust benchmarking for treatment-response prediction, segmentation, and downstream radiomics. A rigorous quality-control workflow—including expert visual assessment and inter-rater reliability analyses—creates a reliable gold-standard resource while highlighting the utility of binary quality categorization for scalability. The dataset supports image synthesis, image standardization, and foundational-model fine-tuning efforts (e.g., MedSAM/SAM), with practical impact on breast cancer diagnostics, treatment response forecasting, and personalized therapy.

Abstract

Artificial Intelligence (AI) research in breast cancer Magnetic Resonance Imaging (MRI) faces challenges due to limited expert-labeled segmentations. To address this, we present a multicenter dataset of 1506 pre-treatment T1-weighted dynamic contrast-enhanced MRI cases, including expert annotations of primary tumors and non-mass-enhanced regions. The dataset integrates imaging data from four collections in The Cancer Imaging Archive (TCIA), where only 163 cases with expert segmentations were initially available. To facilitate the annotation process, a deep learning model was trained to produce preliminary segmentations for the remaining cases. These were subsequently corrected and verified by 16 breast cancer experts (averaging 9 years of experience), creating a fully annotated dataset. Additionally, the dataset includes 49 harmonized clinical and demographic variables, as well as pre-trained weights for a baseline nnU-Net model trained on the annotated data. This resource addresses a critical gap in publicly available breast cancer datasets, enabling the development, validation, and benchmarking of advanced deep learning models, thus driving progress in breast cancer diagnostics, treatment response prediction, and personalized care.

A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations

TL;DR

This work delivers MAMA-MIA, the largest public multicenter breast cancer DCE-MRI benchmark with expert segmentations (1506 pre-treatment cases) by harmonizing four TCIA collections and generating 1506 expert masks through correction of preliminary nnU-Net segmentations. It provides 3D tumor and non-m mass-enhanced region annotations, 49 harmonized clinical/demographic variables, and pretrained baseline nnU-Net weights, enabling robust benchmarking for treatment-response prediction, segmentation, and downstream radiomics. A rigorous quality-control workflow—including expert visual assessment and inter-rater reliability analyses—creates a reliable gold-standard resource while highlighting the utility of binary quality categorization for scalability. The dataset supports image synthesis, image standardization, and foundational-model fine-tuning efforts (e.g., MedSAM/SAM), with practical impact on breast cancer diagnostics, treatment response forecasting, and personalized therapy.

Abstract

Artificial Intelligence (AI) research in breast cancer Magnetic Resonance Imaging (MRI) faces challenges due to limited expert-labeled segmentations. To address this, we present a multicenter dataset of 1506 pre-treatment T1-weighted dynamic contrast-enhanced MRI cases, including expert annotations of primary tumors and non-mass-enhanced regions. The dataset integrates imaging data from four collections in The Cancer Imaging Archive (TCIA), where only 163 cases with expert segmentations were initially available. To facilitate the annotation process, a deep learning model was trained to produce preliminary segmentations for the remaining cases. These were subsequently corrected and verified by 16 breast cancer experts (averaging 9 years of experience), creating a fully annotated dataset. Additionally, the dataset includes 49 harmonized clinical and demographic variables, as well as pre-trained weights for a baseline nnU-Net model trained on the annotated data. This resource addresses a critical gap in publicly available breast cancer datasets, enabling the development, validation, and benchmarking of advanced deep learning models, thus driving progress in breast cancer diagnostics, treatment response prediction, and personalized care.
Paper Structure (6 sections, 7 figures, 3 tables)

This paper contains 6 sections, 7 figures, 3 tables.

Table of Contents

  1. Treatment Response and Survival Prediction.
  2. Automatic Segmentation of Breast Cancer in MRI.
  3. Segmentation Quality Control.
  4. Image Synthesis.
  5. Image Standardization.
  6. Fine-tuning of Foundational Models.

Figures (7)

  • Figure 1: Summary of the main contributions in the MAMA-MIA dataset. The dataset includes three tables with the harmonized clinical and imaging data, train and test splits for benchmarking, the automatic segmentation quality scores, the images and the expert and the automatic preliminary segmentations. Each case in the dataset consists of a pre-treatment T1-weighted DCE-MRI sequence with all the phases in a subfolder under the images folder, and two primary tumour segmentations, one expert-corrected and the preliminary automatic segmentation, without expert corrections.
  • Figure 2: Selection criteria to build the MAMA-MIA dataset. The DCE-MRI cases are collected from four public collections available on TCIA 8: level 2b cohort 24 from I-SPY1/ACRIN 6657 trial (I-SPY1) 9, I-SPY2/ACRIN 6698 trial 2526, NACT-Pilot 27, and Duke-Breast-Cancer-MRI 28, referred to as NACT, ISPY1, ISPY2, and DUKE, respectively.
  • Figure 3: Pre-treatment DCE-MRI sequences from the four collections forming the dataset. Left to right: images are shown in the acquisition plane (axial or sagittal) from DUKE, ISPY1, ISPY2 and NACT. Only two phases of the DCE-MRI sequence are shown, the pre-contrast phase (first row) and the first post-contrast phase (second row).
  • Figure 4: Volumetric analysis image with the corresponding tumor bounding box (in purple) and the tumor volume extracted using the signal enhancement ratio (SER) method denoted as Functional Tumor Volume (FTV). In comparison to FTV, the expert segmentation of the tumor is more precise and contains only the malignant tissues.
  • Figure 5: Manual segmentations were performed for both a) primary tumors and b) non-mass enhanced areas. For each case, the middle slice of the manual segmentation is displayed in sagittal, coronal, and axial views, with the segmentation contour highlighted in red. The rightmost columns in a) and b) present the corresponding 3D segmentation. The images from the first columns correspond to different examples of primary tumor segmentations, meanwhile, the images of the second cases show the more challenging segmentation of non-mass enhanced cancers.
  • ...and 2 more figures