---
title: 'BreastDCEDL_AMBL: DCE-MRI Benchmark'
url: https://www.emergentmind.com/topics/breastdcedl_ambl
type: topic
---

# BreastDCEDL_AMBL: DCE-MRI Benchmark

BreastDCEDL_AMBL is a curated benchmark dataset for automated classification of breast lesions in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), derived from The Cancer Imaging Archive’s AMBL collection and standardized for deep learning research. In its benchmark form, it was curated into a dataset with 88 patients and 133 annotated lesions, including 89 benign and 44 malignant lesions, with public release of the dataset, models, and evaluation protocols. The same name also appears in external validation work on HER2 prediction, where BreastDCEDL_AMBL is described as an independent, single-center dataset comprising 43 lesions from 42 patients with segmented tumor masks and HER2 status annotations. Across these usages, the term denotes a standardized AMBL-based resource for reproducible breast DCE-MRI modeling, especially in settings where benign lesion annotation is essential and where cross-institutional generalization is under study [2509.26440][2510.13897].

## 1. Origin and intended role

BreastDCEDL_AMBL was introduced to address a specific infrastructure deficit in public breast MRI resources: prior public datasets lacked comprehensive benign lesion annotations, which limited benign-versus-malignant classification research. The benchmark dataset was created from the AMBL collection specifically to enable lesion-level DCE-MRI classification with standardized splits and protocols. In this formulation, it is presented as the first public DCE-MRI resource with full, expert-verified benign and malignant lesion annotations and raw data integrity, and as the first standardized benchmark dataset for lesion-level DCE-MRI classification [2509.26440].

The dataset is also connected to the broader BreastDCEDL program. BreastDCEDL, introduced as a curated, deep learning-ready dataset of pre-treatment 3D DCE-MRI from 2,070 breast cancer patients drawn from I-SPY1, I-SPY2, and Duke, provides the large-scale multicenter substrate for model development, while BreastDCEDL_AMBL supplies a lesion-annotation regime that includes benign findings and therefore supports lesion classification tasks not directly available in the primary BreastDCEDL cohorts [2506.12190].

## 2. Curation pipeline and representation

The benchmark curation of BreastDCEDL_AMBL began from the original AMBL collection of 632 DCE-MRI exams. From this source, 88 patients with complete T1-weighted DCE series consisting of 1 pre-contrast and 4 post-contrast acquisitions, together with comprehensive manual lesion segmentations, were selected. Raw DICOM slices were reorganized into 3D NIfTI volumes while preserving native signal intensities [2509.26440].

For each patient, the dataset includes a comprehensive lesion mask with unique integer values representing each lesion, and a malignant-only mask for primary tumor regions. Lesion definition includes both masses and non-mass enhancement, following BI-RADS terminology. Stratified sampling was used so that training, validation, and test splits maintain proportional representation of malignant and benign lesions as well as HER2 status [2509.26440].

This curation philosophy is consistent with the broader BreastDCEDL effort, which standardized raw DICOM imaging into high-fidelity 3D NIfTI volumes with preserved original intensity and spatial information, unified tumor annotations, and harmonized clinical metadata including pathologic complete response, hormone receptor status, and HER2 status. A plausible implication is that BreastDCEDL_AMBL extends this standardization logic to an AMBL-derived lesion-classification setting in which benign lesion masks are indispensable [2506.12190].

## 3. Cohort composition and label structure

In its lesion-classification benchmark form, BreastDCEDL_AMBL contains 88 patients and 133 annotated lesions, of which 44 are malignant and 89 are benign. The reported split structure is as follows [2509.26440]:

| Subset | Patients | Total lesions |
|---|---:|---:|
| Training | 58 | 88 |
| Validation | 15 | 22 |
| Test | 15 | 23 |

The lesion-type distribution across the full cohort is 29 malignant and 59 benign lesions in training, 7 malignant and 15 benign lesions in validation, and 8 malignant and 15 benign lesions in test. At the patient level, 31 patients present malignant tumors, 42 have only benign findings, and 15 have both lesion types [2509.26440].

A separate description appears in external HER2-validation work, where BreastDCEDL_AMBL is characterized as an independent, single-center dataset comprising 43 lesions from 42 patients, with segmented tumor masks and HER2 status annotations, collected under different imaging vendors, protocols, and patient population relative to the multicenter I-SPY/BreastDCEDL cohort [2510.13897]. This suggests a task-specific subset or alternate curation view of AMBL-derived data, although the publications do not explicitly reconcile the two cohort definitions.

## 4. Benchmark methodology and transformer baseline

For the lesion-classification benchmark, a transformer-based framework centered on SegFormer was implemented after comparison with Vision Transformer, SETR, Swin Transformer, and TransUNet. The model operates on RGB-fused \(256 \times 256\) patches in which the pre-contrast, first post-contrast, and last post-contrast images are mapped to channels. Its output is a spatial probability mask of malignancy, so classification is derived from semantic segmentation rather than from a direct image-level classifier [2509.26440].

Training used a hybrid objective,
\[
\mathcal{L}_{\text{combined}} = \mathcal{L}_{\text{BCE}} + \mathcal{L}_{\text{Dice}},
\]
with
\[
\mathcal{L}_{\text{BCE}} = -\frac{1}{N} \sum_{i=1}^{N} [t_i \log(p_i) + (1-t_i)\log(1-p_i)]
\]
and
\[
\mathcal{L}_{\text{Dice}} = 1 - \frac{2\sum_{i} p_i t_i + \varepsilon}{\sum_{i} p_i + \sum_{i} t_i + \varepsilon}.
\]
Lesion-level malignancy was determined by a pixel-ratio rule,
\[
\text{Malignancy Score} = \frac{\# \text{malignant pixels}}{\# \text{total lesion pixels}},
\]
with lesions classified as malignant when the score was at least \(0.3\) [2509.26440].

Because AMBL is comparatively small, training incorporated an expanded cohort of over 1,200 patients through integration with BreastDCEDL-ISPY1 and BreastDCEDL-ISPY2. These auxiliary datasets contain only primary malignant tumor masks, but were used to pre-train lesion-centered patch models before fine-tuning for benign-versus-malignant discrimination on BreastDCEDL_AMBL. This transfer-learning design explicitly leverages the scale of BreastDCEDL while reserving BreastDCEDL_AMBL for the benchmark task that depends on benign annotations [2509.26440][2506.12190].

## 5. Reported performance and interpretability

The SegFormer baseline achieved a lesion-level test AUC of 0.92 and a lesion-level test accuracy of 82%. For malignant segmentation on the test set, the reported mean Dice coefficient is 0.54. At the patient level, the model achieved 100% sensitivity and 67% specificity, with the associated claim that one-third of unnecessary biopsies could be avoided without missing malignancies [2509.26440].

Interpretability is built into the benchmark through pixel-wise malignancy probability maps and binary segmentations thresholded from these maps. The framework therefore produces spatial predictions indicating where malignant features are estimated to lie within a lesion. Visual overlays of ground truth, predictions, and their overlap are part of the reported evaluation. In multi-lesion patients, each lesion is processed independently, which aligns the benchmark with the clinical reality of multiple suspicious findings in a single examination [2509.26440].

These design choices distinguish BreastDCEDL_AMBL from DCE-MRI resources that only provide primary tumor masks. In the benchmark paper, this distinction is not merely organizational: it is the condition that makes benign-versus-malignant lesion classification reproducibly measurable in a public setting [2509.26440].

## 6. External validation use in HER2 prediction

BreastDCEDL_AMBL also serves as an external validation set for noninvasive HER2 prediction from DCE-MRI. In that context, a Triple-Head Dual-Attention ResNet was trained solely on the multicenter I-SPY/BreastDCEDL cohort and evaluated on BreastDCEDL_AMBL without training or fine-tuning. Preprocessing, inference, and prediction aggregation on AMBL followed the exact same protocols used for I-SPY, and patient-level HER2 predictions were obtained by aggregating slice-level outputs, with median aggregation reported as optimal for AUC [2510.13897].

The validation study compared seven normalization and clipping strategies. The best external AMBL performance was obtained with global min-max normalization, yielding accuracy \(0.674\), AUC \(0.664\), sensitivity \(0.56\), and specificity \(0.74\). Channel upper clipping at \(q_{0.98}\) produced accuracy \(0.651\), AUC \(0.613\), sensitivity \(0.69\), and specificity \(0.63\). Per-slice min-max normalization yielded accuracy \(0.651\) and AUC \(0.655\). Across all reported normalization schemes, AUC remained above \(0.60\), which the study interprets as consistent discriminative ability under domain shift [2510.13897].

The same study reports that N4 bias field correction slightly degraded performance and was omitted from final AMBL evaluation. It also states that transformer models were not directly tested on AMBL; their weaker performance was established on I-SPY test data rather than on the external AMBL cohort [2510.13897].

## 7. Position within breast DCE-MRI research

BreastDCEDL_AMBL occupies a specific niche within breast DCE-MRI datasets. BreastDCEDL provides multicenter scale, standardized NIfTI conversion, unified tumor annotations, harmonized metadata, and predefined benchmark splits for clinically meaningful tasks such as treatment response prediction and biomarker modeling. BreastDCEDL_AMBL contributes the missing lesion-level benign annotation layer needed for benign-versus-malignant classification, while also functioning as a cross-institutional test bed for transferability of models trained elsewhere [2506.12190][2509.26440][2510.13897].

Two points of interpretation are important. First, the coexistence of an 88-patient, 133-lesion benchmark description and a 43-lesion, 42-patient external-validation description indicates that “BreastDCEDL_AMBL” is used in more than one operational sense. A plausible implication is that one publication formalizes a comprehensive lesion-classification benchmark, whereas another uses a HER2-labeled external subset or task-specific extraction. Second, the dataset’s significance stems not only from standardization but from the combination of standardized imaging with full benign and malignant lesion masks, which had been absent in prior public DCE-MRI resources [2509.26440][2510.13897].

In this sense, BreastDCEDL_AMBL is best understood as an AMBL-derived, deep learning-oriented benchmark layer within a broader ecosystem of standardized breast MRI resources. Its main technical value lies in making lesion-level DCE-MRI classification, transfer learning from larger tumor-only cohorts, and external validation under acquisition shift all experimentally tractable within a public and reproducible framework [2509.26440][2506.12190].

Source: https://www.emergentmind.com/topics/breastdcedl_ambl