LoGo-MR: Screening Breast MRI for Cancer Risk Prediction by Efficient Omni-Slice Modeling

Xin Wang; Yuan Gao; George Yiasemis; Antonio Portaluri; Zahra Aghdam; Muzhen He; Luyi Han; Yaofei Duan

cs.CVApr 13, 2026

LoGo-MR: Screening Breast MRI for Cancer Risk Prediction by Efficient Omni-Slice Modeling

Authors: Xin Wang, Yuan Gao, George Yiasemis, Antonio Portaluri, Zahra Aghdam, Muzhen He, Luyi Han, Yaofei Duan, +8 more

Organizations: Netherlands Cancer Institute · Radboud University Medical Center · University Medical Center Utrecht · Macao Polytechnic University

Abstract

Efficient and explainable breast cancer (BC) risk prediction is critical for large-scale population-based screening. Breast MRI provides functional information for personalized risk assessment. Yet effective modeling remains challenging as fully 3D CNNs capture volumetric context at high computational cost, whereas lightweight 2D CNNs fail to model inter-slice continuity. Importantly, breast MRI modeling for shor- and long-term BC risk stratification remains underexplored. In this study, we propose LoGo-MR, a 2.5D local-global structural modeling framework for five-year BC risk prediction. Aligned with clinical interpretation, our framework first employs neighbor-slice encoding to capture subtle local cues linked to short-term risk. It then integrates transformer-enhanced multiple-instance learning (MIL) to model distributed global patterns related to long-term risk and provide interpretable slice importance. We further apply this framework across axial, sagittal, and coronal planes as LoGo3-MR to capture complementary volumetric information. This multi-plane formulation enables voxel-level risk saliency mapping, which may assist radiologists in localizing risk-relevant regions during breast MRI interpretation. Evaluated on a large breast MRI screening cohort (~7.5K), our method outperforms 2D/3D baselines and existing SOTA MIL methods, achieving AUCs of 0.77-0.69 for 1- to 5-year prediction and improving C-index by ~6% over 3D CNNs. LoGo3-MR further improves overall performance with interpretable localization across three planes, and validation across seven backbones shows consistent gains. These results highlight the clinical potential of efficient MRI-based BC risk stratification for large-scale screening. Code will be released publicly.

Figures & tables

Figure 1: Comparison of 3D, 2D, and the proposed LoGo-MR and LoGo 3 -MR frameworks for BC risk prediction. LoGo-MR combines local slice fusion with global MIL-based sequence modeling to capture both short-range anatomical continuity and long-range slice dependencies, while providing interpretable slice-level importance. LoGo 3 -MR extends this strategy across axial, sagittal, and coronal planes to capture complementary anatomical cues.

	Method	FLOPs	FPS	1Y AUC	2Y AUC	3Y AUC	4Y AUC	5Y AUC	Mean AUC	C-index
ResNet18	3D Baseline	1170	4	$56.8$ $\pm$ 8.1	$58.1$ $\pm$ 6.6	$56.7$ $\pm$ 5.6	$55.1$ $\pm$ 5.1	$55.6$ $\pm$ 4.7	$56.5$ $\pm$ 5.4	$56.4$ $\pm$ 4.6
	3D MFFN [ 9 ]	1187	3	$62.4$ $\pm$ 7.5	$60.6$ $\pm$ 6.3	$58.9$ $\pm$ 5.5	$56.8$ $\pm$ 4.9	$56.6$ $\pm$ 4.7	$59.1$ $\pm$ 5.1	$56.7$ $\pm$ 4.6
	2D Baseline (Mean)	354	18	$63.7$ $\pm$ 9.3	$61.1$ $\pm$ 7.2	$60$ $\pm$ 5.7	$55.6$ $\pm$ 5.3	$54.7$ $\pm$ 5.1	$59$ $\pm$ 5.9	$56.2$ $\pm$ 4.9
	2D MoE [ 16 ]	354	17	$61.4$ $\pm$ 7.9	$59.6$ $\pm$ 6.7	$57.9$ $\pm$ 5.6	$59$ $\pm$ 5.0	$59.9$ $\pm$ 4.8	$59.6$ $\pm$ 5.3	$57.6$ $\pm$ 4.6
	2D LSTM-MIL [ 20 ]	354	17	$67.8$ $\pm$ 8.0	$64.6$ $\pm$ 6.5	$61.6$ $\pm$ 5.5	$58.5$ $\pm$ 5.1	$57.5$ $\pm$ 4.8	$62$ $\pm$ 5.4	$58.8$ $\pm$ 4.7
	2D MambaMIL [ 19 ]	354	17	$68.3$ $\pm$ 7.7	$60.2$ $\pm$ 6.8	$58$ $\pm$ 5.9	$57.3$ $\pm$ 5.2	$56.8$ $\pm$ 5.0	$60.1$ $\pm$ 5.4	$56.2$ $\pm$ 4.8

Table 2: Performance comparison of different methods across prediction horizons.

Figure 3: Comparison of 3D CNNs, 2D baselines, and the proposed LoGo-MR-RISK and LoGo 3 -MR-RISK across seven backbones.

Figure 4: Multi-plane risk localization with LoGo 3 -MR. For each case, we present: (Left) Slice-level importance weights learned along the axial (Z), coronal (Y), and sagittal (X) planes. (Right) The maximum intensity projection (MIP) view and corresponding high-importance regions onto orthogonal views, illustrating spatially consistent risk-relevant areas localization across planes over time.

	Lo	Go	Method	1Y AUC	2Y AUC	3Y AUC	4Y AUC	5Y AUC	Mean AUC	C-index
ResNet18	$\times$	$\times$	2D Baseline: $g=0$	$63.7$ $\pm$ 9.3	$61.1$ $\pm$ 7.2	$60$ $\pm$ 5.7	$55.6$ $\pm$ 5.3	$54.7$ $\pm$ 5.1	$59$ $\pm$ 5.9	$56.2$ $\pm$ 4.9
	✓	$\times$	Gap: $g=1$	$66$ $\pm$ 9.1	$59.5$ $\pm$ 7.1	$55.8$ $\pm$ 6.1	$53.3$ $\pm$ 5.6	$52.9$ $\pm$ 5.2	$57.5$ $\pm$ 5.9	$55.1$ $\pm$ 5.0
	✓	$\times$	Gap: $g=3$	$68.9$ $\pm$ 8.2	$63.2$ $\pm$ 6.6	$59$ $\pm$ 5.6	$56.8$ $\pm$ 5.1	$56.2$ $\pm$ 4.9	$60.8$ $\pm$ 5.4	$57.5$ $\pm$ 4.7
	✓	$\times$	Gap: $g=5$	$67.8$ $\pm$ 8.4	$63$ $\pm$ 6.8	$60.6$ $\pm$ 5.6	$58.3$ $\pm$ 5.0	$59$ $\pm$ 4.8	$61.8$ $\pm$ 5.4	$59.2$ $\pm$ 4.7
	✓	$\times$	Gap: $g=7$	$63.9$ $\pm$ 8.8	$62$ $\pm$ 6.9	$59.3$ $\pm$ 5.8	$56.1$ $\pm$ 5.2	$54.7$ $\pm$ 5.0	$59.2$ $\pm$ 5.6	$56.8$ $\pm$ 4.8
	$\times$	✓	w/o Lo & w/o Pos	$64.9$ $\pm$ 9.0	$60.2$ $\pm$ 7.4	$59.8$ $\pm$ 6.0	$58.7$ $\pm$ 5.3	$58.7$ $\pm$ 5.0	$60.5$ $\pm$ 5.9	$58.3$ $\pm$ 5.0

Table 3: Ablation study of local and global structural modeling.

Explore similar work

Longitudinal Multi-View Breast Cancer Risk Prediction
Jul 13, 2026Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin +5Breast Cancer Longitudinal Medical Image Analysis
Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, Norway · Norwegian Computing Center, Oslo, Norway +1
Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories
Oct 12, 2026Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin +5Breast Cancer Mammography
Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, 9037, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, 9019, Norway · Norwegian Computing Center, Oslo, 0373, Norway +1
Revealing Mammographic Phenotypes in Deep Learning Breast Cancer Risk Models
Jun 24, 2026Ruiyu Jia, Yanqi Xu, Yuxuan Chen +2Breast Cancer Medical Image Analysis
Department of Mathematics, New York University Shanghai, China · Center for Data Science, New York University, USA · Perlmutter Cancer Center, NYU Langone Health, New York, USA +1

Jul 13, 2026J/K move · Enter open · S save

Longitudinal Multi-View Breast Cancer Risk Prediction

Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin, Kristoffer Wickstrøm+4

Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, Norway · SPKI The Norwegian Centre for Clinical Artificial Intelligence, University Hospital of North Norway, Tromsø, Norway · Norwegian Computing Center, Oslo, Norway+1

Breast Cancer Longitudinal Medical Image Analysis Multi-View Learning Mammography Clinical Risk Prediction

Open paper AskPDF arXiv

Accurate breast cancer risk prediction from screening mammography is critical for enabling personalized screening intervals and early detection. Recent deep learning methods have shown the value of longitudinal data and explicit temporal alignment. However, existing approaches either perform explicit alignment using a single mammographic view or model multiple views without explicit longitudinal alignment, limiting their ability to exploit the complementary spatial-temporal information used in clinical practice. To address this gap, we propose LMV-Net, a longitudinal multi-view breast cancer risk prediction model that jointly analyzes anatomically complementary CC and MLO views within an explicitly aligned longitudinal framework. We evaluate our approach on the public EMBED and CSAW-CC datasets, comparing it to state-of-the-art breast cancer risk prediction methods. Our model consistently outperforms existing approaches in overall risk prediction performance and across different breast density and cancer subgroups. Importantly, these improvements highlight the potential of longitudinal multi-view modeling to enhance risk stratification, paving the way for future work on personalized screening, earlier identification of high-risk patients, and more efficient screening resource allocation. The code is available at https://github.com/sot176/LMV-Net.

Abstract

Explore similar work

Longitudinal Multi-View Breast Cancer Risk Prediction

Explicitly Aligned Multi-View Breast Cancer Risk Prediction from Variable-Length Screening Histories

Revealing Mammographic Phenotypes in Deep Learning Breast Cancer Risk Models