Source-linked AI summary
Evaluating ADC-only deep learning pipelines for breast cancer detection and segmentation using standalone diffusion-weighted MRI
Pablo García Marcos, Paula Puerta Gonzĺez, Guillermo Lorenzo, Héctor Gómez, Covadonga del Camino, Angel Rio-Alvarez, Víctor M. González
TL;DR
DCE-MRI is the breast-cancer imaging gold standard but requires contrast and long acquisitions, while deep-learning evidence for DW-MRI and ADC maps remains limited. This paper evaluates ADC-only approaches across classification, detection, and segmentation; classification and detection show potential but have limited recall, and segmentation metrics remain subpar versus contrast-based imaging.
Problem
Deep-learning research has focused on DCE-MRI, while ADC-only approaches for breast-tumor detection and segmentation remain relatively unexplored despite DW-MRI’s faster, contrast-free acquisition.
Source: §1 · p. 2Method
The study evaluates ADC-only deep-learning approaches for slice classification, object detection, and semantic tumor segmentation.
Source: §1 · p. 2, §4.1.1 · p. 4Results
ADC-based slice classification and object detection showed potential but had limited recall, while segmentation metrics remained subpar versus contrast-based imaging.
Source: §6 · p. 6, §5.3 · p. 6Takeaways & Limitations
ADC maps may aid screening applications for patients with dense breast tissue, although low recall may restrict integration into screening workflows.
Source: §6 · p. 6The multicenter dataset is heterogeneous, and its 4-mm inter-slice gap may particularly degrade 3D semantic segmentation performance.
Source: §3 · p. 3
Abstract
from arXiv · showhide
Dynamic contrast-enhanced (DCE) imaging is the gold standard technique for the detection and characterization of breast cancer using magnetic resonance imaging (MRI). However, DCE-MRI requires long acquisition times and the administration of contrast into the bloodstream, which can cause allergic reactions. Alternatively, diffusion-weighted MRI (DW-MRI) is a standard complementary technique for breast MRI that does not require contrast, has shorter acquisition times, and enables calculation of apparent diffusion coefficient (ADC) maps that correlate with tumor cellularity. Yet, despite these technical advantages, deep learning research has focused on DCE-based models and has barely explored the tumor detection performance of DW-MRI and ADC maps either in combination with DCE-MRI or as standalone alternatives. Here, we evaluate the application of different state-of-the-art deep learning techniques for detection and segmentation of breast cancer using ADC-only images. This is, to our knowledge, the first comprehensive evaluation of ADC-only breast cancer pipelines for classification, detection, and segmentation tasks.
1. Introduction
DCE-MRI is the breast-cancer imaging gold standard but requires contrast and long acquisitions, while ADC-only deep-learning approaches remain relatively unexplored. This study evaluates ADC-based approaches for identifying, localizing, and segmenting breast tumors.
1. Introduction
- Breast cancer is highly heterogeneous, with subtypes that have distinct molecular profiles relevant to treatment strategies and prognosis.The paper describes it as the most diagnosed tumor type and leading cause of cancer-related death among women worldwide.Source: §1 · p. 1
- DCE-MRI is the gold standard for breast-cancer detection and characterization, but requires contrast agents and long, complex acquisitions.Contrast agents may trigger allergic reactions in some patients; multiple time points and post-processing contribute to acquisition burden.Source: §1 · p. 2
- DW-MRI is a faster, non-contrast alternative, and ADC maps derived from it correlate with tumor cellularity and other diagnostic tissue properties.Tumor tissue typically shows restricted water diffusion, appearing enhanced on DW-MRI and low on ADC maps.Source: §1 · p. 2
- Deep-learning models have focused on DCE-MRI, with relatively few efforts using DW-MRI or ADC maps alone or alongside DCE-MRI.The authors identify ADC-only deep-learning limitations as relatively unexplored.Source: §1 · p. 2
- The study evaluates ADC-based object detection for identifying and localizing cancer and CNNs for semantic tumor segmentation.The broader objective is to assess the viability of different deep-learning approaches for breast-cancer identification using ADC imaging.Source: §1 · p. 2
2. Related work
Prior breast-cancer detection and segmentation work has largely used mammography, tomosynthesis, or DCE-MRI, with some hybrid DCE/DW approaches. Direct DW-MRI or ADC lesion detection and segmentation remain limited.
2. Related work
- OD models have been widely used for breast-lesion detection and localization in mammography and tomosynthesis, but DW-MRI or ADC-based OD approaches are scarce.Examples include YOLO-based systems and a two-stage VGG19–YOLOv5 pipeline.Source: §2 · p. 2
- Breast MRI tumor detection and segmentation have traditionally used machine-learning approaches on DCE-MRI, including unsupervised methods, 3D U-Net, and transformer/CNN architectures.Some models also integrate spatial and temporal information from pre- and post-contrast images.Source: §2 · p. 2
- Some studies combine DCE-MRI and DW-MRI for lesion classification or segmentation, while work exclusively using DW-MRI for lesion detection and segmentation remains limited.The cited hybrid approaches include feature-based classification and deep-learning models using both modalities.Source: §2 · pp. 2–3
3. Materials and Data
The study uses pre-treatment ADC data, expert tumor masks, and the ACRIN 6698 multicenter cohort to evaluate ADC-only deep-learning approaches. Dataset heterogeneity and a 4-mm inter-slice gap are noted as relevant limitations, particularly for 3D segmentation.
3. Materials and Data
- ACRIN 6698 includes n=385 breast-cancer patients with up to four MRI examinations at clinical timepoints, original DW-MRI scans, monoexponential ADC maps, and expert-created tumor masks; it was created for a multicenter neoadjuvant-chemotherapy trial evaluating ADC reliability and therapeutic-response prediction.Source: §3 · p. 3
- The multicenter dataset is heterogeneous, and its 4-mm inter-slice gap may degrade performance, especially for 3D semantic segmentation.The authors note that scanner and acquisition variability may reduce performance, while making results more representative of multicenter application.Source: §3 · p. 3
- Figure 1 presents a study slice in DW, ADC, and mask formats.The panels are ordered DW on the left, ADC in the center, and mask on the right.Source: Figure 1 · p. 3
- Images with masks not labeled high quality were removed, then resized to 256×256 and normalized to [0, 1]; 3D volumes were also created for segmentation.For segmentation, images and volumes were used at full resolution without ROI cropping; OD boxes were generated from tumor masks with a 6-voxel margin.Source: §3 · p. 3
4. Methods
The methods evaluate slice classification, object detection, and semantic segmentation using ADC maps with a single b-value. They pair YOLO-based classification and detection with U-Net-based segmentation and task-specific evaluation metrics.
4. Methods
- All three tasks use ADC maps with one single b-value as input.Source: §4 · p. 3
4.1.1. Architecture for slice classification and object detection
- YOLO handles slice classification and object detection; a slice is tumor-free only when no predicted box probability exceeds 0.25.Source: §4.1.1 · p. 4
4.1.2. Architecture for semantic segmentation
- The study uses U-Net for semantic segmentation, then employs nnU-Net on full-resolution images and volumes after ground-truth-guided crops performed poorly and were unsuitable for inference.Source: §4.1.2 · p. 4
4.2. Training, validation and test data. Optimization algorithms
- Five-fold cross-validation holds out 20% of patients for testing in each fold and splits the remainder into three-quarters training and one-quarter validation.Source: §4.2 · p. 4
4.3. Model performance metrics
- Accuracy uses TP and TN over all outcomes; outcomes are defined at slice, box, or voxel level by task, and detection also uses AP, the area under the precision-recall curve.For detection, a true positive requires a predicted box to overlap ground truth by more than 50%; segmentation counts only boxes with confidence of at least 0.25.Source: §4.3 · p. 4
5. Results
The results evaluate ADC-based models for slice classification, tumor detection and localization, and semantic segmentation. Classification and localization show promising performance, while segmentation remains subpar.
5. Results
- The reported results span slice classification, tumor detection and localization, and semantic segmentation.The paper presents performance tables for each task.Source: §5 · p. 4, Table 1 · p. 5, Table 2 · p. 5
5.1. Slice classification
- 94.2% mean precision and 91.0% overall accuracy were achieved in slice classification.The model correctly classified most slices, but challenging cases included a cancerous slice classified as non-cancerous.Source: §5.1 · p. 5, Table 1 · p. 5
5.2. Tumor detection and localization
- 87.8% precision, 68.0% recall, and 75.8% average precision were achieved for tumor detection and localization.The examples include a correctly localized tumor and a false-positive box on non-cancerous tissue.Source: §5.2 · pp. 5–6, Table 2 · p. 5
5.3. Semantic segmentation of cancerous tumors
- 2D and 3D models achieved similar DICE scores, while all networks had subpar segmentation metrics, particularly compared with contrast-based imaging.The authors suggest that the dataset’s approximately 3–5 mm inter-slice distance likely hindered 3D U-Net performance.Source: Table 3 · p. 5, §5.3 · p. 6
6. Conclusion
ADC maps show potential for breast cancer detection, but low recall constrains classification and object detection, while segmentation remains challenging. Large inter-slice spacing limits volumetric continuity, and the authors identify several directions for further study.
6. Conclusion
- Slice classification and object detection showed detection potential, but both had relatively low recall that may restrict screening use.Classification precision and accuracy supported use as an aid in screening applications for patients with dense breast tissue; object detection achieved good lesion localization despite DW-MRI limitations.Source: §6 · p. 6
- Segmentation achieved only moderate DICE scores, and large inter-slice spacing produced highly anisotropic volumes with reduced anatomical continuity.The authors suggest the available volumetric information may be insufficient to fully exploit 3D segmentation architectures.Source: §6 · p. 6
- The authors propose exploring 3D-volume reconstruction, multimodal DW-MRI and ADC inputs, and improved annotation quality.They cite interpolation work, multimodal studies, and partial radiologist annotations as relevant directions or considerations.Source: §6 · p. 6
CRediT authorship contribution statement
The listed authors contributed across writing, visualization, validation, investigation, and data curation, with several also credited for methodology, conceptualization, or project leadership.
CRediT authorship contribution statement
- Paula Puerta González and Pablo García Marcos are credited with writing, visualization, validation, investigation, and data curation.Both are listed for review and editing as well as the original draft.Source: CRediT authorship contribution statement · p. 6
- Guillermo Lorenzo and Hector Gomez are credited with methodology, investigation, and conceptualization as well as writing, visualization, and validation.Both are listed for review and editing and the original draft.Source: CRediT authorship contribution statement · p. 6
- Covadonga del Camino is credited with resources and conceptualization, while Víctor M. González and Angel Rio-Alvarez have broader leadership and project roles.González is listed for project administration and funding acquisition; Rio-Alvarez for project administration and supervision.Source: CRediT authorship contribution statement · p. 6
Ethics statement
The study used publicly available, fully de-identified breast cancer imaging data and reports no identified ethical complications.
Ethics statement
- The study used publicly available, fully de-identified breast cancer imaging data from The Cancer Imaging Archive.The authors state that institutional review board approval and informed consent were not required.Source: Ethics statement · p. 7
- The authors report that no ethical complications were identified during the study.Source: Ethics statement · p. 7