Abstract
Breast magnetic resonance imaging (MRI) protocols usually consist of multiple acquisitions with distinct image contrasts and diagnostic roles. Accurate MRI sequence categorization is therefore necessary for large-scale data curation and artificial intelligence pipelines. This is commonly implemented based on DICOM metadata, which can, however, often be heterogeneous and sometimes even inconsistent or incomplete. In this study, we investigated whether breast MRI sequence types can be reliably recognized directly from image data using convolutional neural networks (CNNs). We utilized acquisitions from the ODELIA, AMBL, and DUKE datasets and manually assigned them to six clinically relevant classes, i.e., T1-weighted, T1-weighted contrast enhanced, T1-weighted fat-suppressed, T1-weighted fat-suppressed and contrast-enhanced, T2-weighted, and T2-weighted fat-suppressed. Three CNN architectures, i.e., ResNet-18, DenseNet-121, and EfficientNet-B0, were trained using the basic ODELIA and the extended ODELIA training datasets independently, the latter including a small subset of AMBL and DUKE cases. Classification performance was evaluated using independent ODELIA, AMBL, and DUKE evaluation datasets. All models achieved high accuracy on ODELIA, with values ranging from 96.1% to 98.1%. External performance was also high on AMBL, but it was substantially lower on DUKE when models were trained only on ODELIA, with error analysis indicating frequent mistakes between T1-weighted and T2-weighted acquisitions, most likely associated with the differences between spin-echo and gradient-echo sequences. However, limited external data present in ODELIA markedly improved model accuracy on DUKE from 58.2–60.5% to 92.2–94.4%. Our findings demonstrate that exclusively image-based breast MRI sequence type classification using CNNs is feasible, but robust generalization requires exposure to diverse examination protocols.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_035.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=0VWR6qY0EP
BibTex
@InProceedings{GrzMat_Breast_MICCAISAT2026,
author = { Grzyb, Mateusz AND Liebert, Andrzej},
title = { { Breast MRI Sequence Type Classification using Convolutional Neural Networks } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}