PRISM-Net: Patient-specific reference-guided inter-breast symmetry matching for three-class breast DCE-MRI classification

Abstract Breast DCE-MRI classification of breast sides as no lesion, benign, or malignant is complicated by substantial patient-specific variation in breast anatomy, tissue composition, and physiological enhancement. Although radiologists routinely use the contralateral breast as an internal reference, computational bilateral comparison remains challenging because corresponding contralateral regions are not anatomically identical and bilateral differences are not necessarily pathological. We propose PRISM-Net, a registration-free bilateral framework that treats the contralateral breast as an approximate patient-specific reference. PRISM-Net learns complementary target-appearance and reference-conditioned deviation representations. Adaptive patch-level matching constructs content-dependent contralateral references in a shared feature space without voxel-wise registration, while local-contrast reweighting emphasizes focal deviations over diffuse bilateral mismatch. The resulting deviation features are integrated with intrinsic target-side appearance through residual cross-attention and aggregated across slices for breast-side classification. On ODELIA, PRISM-Net achieved Macro AUC, Micro AUC, and quadratic weighted kappa values of 84.11 ± 2.33, 90.64 ± 1.61, and 60.94 ± 5.64 on the in-distribution test set, and 68.51 ± 4.54, 80.74 ± 2.68, and 43.45 ± 7.10 on the held-out institution, respectively, yielding the highest point estimates among the evaluated methods across the primary metrics. PRISM-Net also demonstrated favorable performance on an independent institutional cohort and background-complexity subsets, while ablation experiments supported the contributions of adaptive reference matching and focal deviation modeling. These findings support patient-specific contralateral reference learning as a clinically grounded strategy for breast DCE-MRI classification under heterogeneous background conditions.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11425380/v1
Primary Topic
MRI in cancer diagnosis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

PRISM-Net: Patient-specific reference-guided inter-breast symmetry matching for three-class breast DCE-MRI classification

Mingxu Wang, Xiru Li, Di Kong, Wenbiao Du et al.
Research Square
MRI in cancer diagnosis
preprint

PRISM-Net: Patient-specific reference-guided inter-breast symmetry matching for three-class breast DCE-MRI classification

Mingxu Wang, Xiru Li, Di Kong, Wenbiao Du, Yiman Zhong, Xiawei Yue, Yuexin Duan, Shuaiwen Zhou, Boya Zhang, Liuquan Cheng
preprint en

Abstract

Abstract Breast DCE-MRI classification of breast sides as no lesion, benign, or malignant is complicated by substantial patient-specific variation in breast anatomy, tissue composition, and physiological enhancement. Although radiologists routinely use the contralateral breast as an internal reference, computational bilateral comparison remains challenging because corresponding contralateral regions are not anatomically identical and bilateral differences are not necessarily pathological. We propose PRISM-Net, a registration-free bilateral framework that treats the contralateral breast as an approximate patient-specific reference. PRISM-Net learns complementary target-appearance and reference-conditioned deviation representations. Adaptive patch-level matching constructs content-dependent contralateral references in a shared feature space without voxel-wise registration, while local-contrast reweighting emphasizes focal deviations over diffuse bilateral mismatch. The resulting deviation features are integrated with intrinsic target-side appearance through residual cross-attention and aggregated across slices for breast-side classification. On ODELIA, PRISM-Net achieved Macro AUC, Micro AUC, and quadratic weighted kappa values of 84.11 ± 2.33, 90.64 ± 1.61, and 60.94 ± 5.64 on the in-distribution test set, and 68.51 ± 4.54, 80.74 ± 2.68, and 43.45 ± 7.10 on the held-out institution, respectively, yielding the highest point estimates among the evaluated methods across the primary metrics. PRISM-Net also demonstrated favorable performance on an independent institutional cohort and background-complexity subsets, while ablation experiments supported the contributions of adaptive reference matching and focal deviation modeling. These findings support patient-specific contralateral reference learning as a clinically grounded strategy for breast DCE-MRI classification under heterogeneous background conditions.

Research Square
Beijing Institute of Technology (CN), Beijing University of Posts and Telecommunications (CN), Nankai University (CN), Chinese PLA General Hospital (CN), Beihang University (CN), Tsinghua University (CN)
Good health and well-being
MRI in cancer diagnosis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon