Transformer Classification of Breast Lesions: The BreastDCEDL_AMBL Benchmark Dataset and 0.92 AUC Baseline

TL;DR

This work tackles the specificity gap in breast DCE-MRI by developing a transformer-based lesion classifier using RGB-fused temporal MRI patches. Leveraging SegFormer and a large-scale, publicly available training corpus augmented with ISPY data, the authors achieve a lesion-level AUC of $0.92$ and 100% patient-level sensitivity, while providing interpretable malignant segmentation maps. A publicly released benchmark, BreastDCEDL_AMBL, together with trained models and evaluation protocols, offers a reproducible platform to benchmark future methods and accelerate clinical translation. The study also delineates clear directions for advancing toward end-to-end detection, volumetric processing, and privacy-preserving multi-institution learning.

Abstract

Breast magnetic resonance imaging is a critical tool for cancer detection and treatment planning, but its clinical utility is hindered by poor specificity, leading to high false-positive rates and unnecessary biopsies. This study introduces a transformer-based framework for automated classification of breast lesions in dynamic contrast-enhanced MRI, addressing the challenge of distinguishing benign from malignant findings. We implemented a SegFormer architecture that achieved an AUC of 0.92 for lesion-level classification, with 100% sensitivity and 67% specificity at the patient level - potentially eliminating one-third of unnecessary biopsies without missing malignancies. The model quantifies malignant pixel distribution via semantic segmentation, producing interpretable spatial predictions that support clinical decision-making. To establish reproducible benchmarks, we curated BreastDCEDL_AMBL by transforming The Cancer Imaging Archive's AMBL collection into a standardized deep learning dataset with 88 patients and 133 annotated lesions (89 benign, 44 malignant). This resource addresses a key infrastructure gap, as existing public datasets lack benign lesion annotations, limiting benign-malignant classification research. Training incorporated an expanded cohort of over 1,200 patients through integration with BreastDCEDL datasets, validating transfer learning approaches despite primary tumor-only annotations. Public release of the dataset, models, and evaluation protocols provides the first standardized benchmark for DCE-MRI lesion classification, enabling methodological advancement toward clinical deployment.

Figures (8)

• Figure 1: Representative lesion segmentation examples from the AMBL dataset showing three patients (rows a--c). Each row displays: peak enhancement phase (column 1), color-coded segmentation with benign (green/blue) and malignant (red) lesions (column 2), segmentation overlay on peak enhancement (column 3), and RGB fusion showing temporal enhancement kinetics (column 4). Patient (a) shows bilateral lesions with one benign and one malignant. Patient (b) demonstrates bilateral malignant lesions. Patient (c) presents bilateral benign lesions. The RGB fusion highlights contrast uptake and washout patterns for lesion characterization.
• Figure 2: Representative lesion segmentation examples from BreastDCEDL datasets. Three patient cases are shown: rows (a) and (b) from BreastDCEDL_ISPY2, and row (c) from BreastDCEDL_ISPY1. Column 1 displays the peak enhancement phase following contrast administration. Column 2 shows the binary tumor segmentation mask. Column 3 presents the segmentation overlay on the peak enhancement image. Column 4 illustrates the RGB fusion combining temporal DCE-MRI phases, where pre-contrast, early post-contrast, and late post-contrast acquisitions are mapped to red, green, and blue channels respectively, enabling visualization of enhancement dynamics.
• Figure 3: SegFormer architecture for breast lesion segmentation. The model consists of a hierarchical transformer encoder with four stages that progressively reduce spatial resolution from $H/4 \times W/4$ to $H/32 \times W/32$ while extracting multi-scale features ($F_1$--$F_4$). Stage 1 employs patch embedding for initial feature extraction, while Stages 2--4 utilize Mix-FFN blocks that combine efficient self-attention with $3 \times 3$ depthwise convolutions for implicit positional encoding. The lightweight All-MLP decoder aggregates these multi-scale features before upsampling by a factor of 4 to produce the final segmentation mask at original resolution.
• Figure 4: Representative examples from a validation batch used during training, showing RGB-fused DCE-MRI slices (left) and their corresponding segmentation masks (right). Approximately half of the masks are empty (all-zero), corresponding to benign lesions in the BreastDCEDL_AMBL dataset where both benign and malignant lesions are fully annotated.
• Figure 5: Examples from a training batch combining all three datasets (BreastDCEDL_ISPY1, BreastDCEDL_ISPY2, and BreastDCEDL_AMBL) with full augmentation applied. The images demonstrate varied lesion appearances and the effects of spatial and photometric augmentations, including rotation, scaling, and intensity adjustments.