CBIS-DDSM: Curated Mammography Dataset
- CBIS-DDSM is a curated mammography dataset derived from DDSM, featuring detailed lesion annotations and comprehensive metadata.
- It supports a wide range of research tasks, including whole-image and ROI classification, segmentation, and localization.
- The dataset exposes challenges such as annotation ambiguity, split inconsistencies, and preprocessing variability in mammography studies.
CBIS-DDSM, the Curated Breast Imaging Subset of the Digital Database for Screening Mammography, is a curated mammography resource derived from DDSM and used as a public benchmark for breast-image analysis. In the literature it is not treated as a single fixed task; rather, it functions as a metadata-rich corpus from which investigators construct breast-level, image-level, ROI-level, and pixel-level problems, often using different subsets, label remappings, and split conventions. That flexibility has made it central to studies of classification, localization, segmentation, multimodal fusion, multi-view modeling, weak supervision, explainability, and dataset harmonization (Wimmer et al., 2021, Zafari et al., 4 Nov 2025).
1. Historical position and dataset identity
Publications consistently describe CBIS-DDSM as the curated version of DDSM, introduced to improve the usability of a legacy screen-film mammography collection for machine learning and computer-aided diagnosis. One line of work characterizes DDSM as a dataset of 2,620 screening exams with four images per exam and describes CBIS-DDSM as the curated lesion subset used to recover cleaner lesion labels and an official train/test partition for downstream modeling (Wimmer et al., 2021). MammoClean similarly presents CBIS-DDSM as a curated subset of DDSM designed to address weaknesses such as imprecise lesion annotations, and summarizes it as a United States screen-film mammography dataset with 1,391 studies and 2,844 images (Zafari et al., 4 Nov 2025).
The corpus is described with nonuniform global counts across papers. Some studies report CBIS-DDSM in terms of findings-only images or task-specific filtered subsets rather than a single canonical size. For example, the Faster R-CNN noise study uses a mass-detection formulation and describes CBIS-DDSM as including only images with findings, totaling 3,089 images, after exclusion of normal exams from the original DDSM framing (Famouri et al., 2021). A ViT-GNN classification paper instead reports 10,239 mammographic images for a normal-versus-abnormal setup (Cai et al., 11 Jul 2025). A multimodal VQA benchmark describes CBIS-DDSM as a database of 2,620 scanned film mammography studies with normal, benign, and malignant cases (Li et al., 15 Aug 2025). This suggests that “CBIS-DDSM” in published experiments often refers simultaneously to the named resource and to task-conditioned reconstructions of it.
The dataset’s importance is amplified by its role as a bridge between classic screen-film mammography and contemporary deep learning. Several papers use it precisely because it exposes methodological issues that do not disappear under curation: split leakage, laterality inconsistencies, class-remapping ambiguity, and strong dependence of performance on whether the task is whole-image, multi-view, or lesion-centered (Zafari et al., 4 Nov 2025, Petrini et al., 2021).
2. Annotation structure and label semantics
CBIS-DDSM is repeatedly described as richer than a bare image archive. Across the cited studies, it is associated with lesion locations, ROI crops, segmentation masks or contours, pathology labels, BI-RADS information, breast density, lesion-type descriptors, and view/laterality metadata, although the exact fields consumed differ by paper. MammoClean lists diagnosis, mass and calcification characteristics, BI-RADS, breast density, laterality, view, and contour-style finding annotations in its harmonized schema, while explicitly noting that age is unavailable in CBIS-DDSM even though it existed in DDSM (Zafari et al., 4 Nov 2025). The multi-task fusion pipeline similarly relies on breast density, lesion type, pathology, and pixel-level lesion annotations derived from the DDSM/CBIS-DDSM combination (Wimmer et al., 2021).
Pathology labels are frequently remapped. In calcification-patch analysis, the three-class label space is malignant, benign, and benign without callback, while a second task collapses malignant and benign into a clinically motivated “Follow-up” class versus benign without callback as “No follow-up” (Panambur et al., 2022). In two-view EfficientNet modeling, the final target is malignant versus non-malignant for one breast side, and exams containing both benign and malignant findings are treated as malignant (Petrini et al., 2021). In multimodal BI-RADS descriptor fusion, the reported experiments focus on mass lesions and benign-versus-malignant classification, even though the underlying dataset also contains calcifications and descriptor vocabularies for both lesion types (Ben-Artzi et al., 2024).
Lesion-type usage is similarly task-dependent. Some papers operate on both mass and calcification subsets, such as ROI fusion with handcrafted features (Tschuchnig et al., 26 Jul 2025), weakly supervised segmentation from classification explanations (Ma et al., 6 Aug 2025), and DSU-Net ROI segmentation built from both mass_case_description_train_set.csv and calc_case_description_train_set.csv (Bozorgpour et al., 3 Jun 2026). Others deliberately restrict the corpus: mass-only object detection with clean segmentation-derived boxes (Famouri et al., 2021), mass-only multi-view classification in Mammo-Mamba (Bayatmakou et al., 23 Jul 2025), or custom micro-mass segmentation benchmarks emphasizing lesions smaller than 200 pixels (Kamran et al., 2022). Conversely, some papers acknowledge that exact class semantics are underreported; the multimodal privileged distillation study states only that image and privileged metadata share a breast-cancer classification label, while the precise label definition for the mass and calcification tasks is not fully specified in the manuscript (Baur et al., 6 Aug 2025). A plausible implication is that CBIS-DDSM is best understood not as a single label ontology but as a source dataset whose annotations are selectively projected into task-specific targets.
3. Experimental roles across the literature
CBIS-DDSM supports a wide spectrum of benchmark formulations. In practice, the same named dataset appears in at least six distinct methodological roles: whole-mammogram or breast-level classification, lesion-centered ROI classification, supervised or weakly supervised segmentation, lesion detection/localization, multimodal metadata fusion, and evaluation-only prompting or federated-learning extensions.
| Task family | Representative formulation | Example papers |
|---|---|---|
| Classification | Whole-image, breast-level, or multi-view benign/malignant or normal/abnormal prediction | (Sarker et al., 24 Feb 2025, Petrini et al., 2021, Cai et al., 11 Jul 2025, Bayatmakou et al., 23 Jul 2025) |
| ROI-level analysis | Benign–malignant classification on cropped lesion ROIs | (Asad et al., 14 Apr 2026, Tschuchnig et al., 26 Jul 2025, Panambur et al., 2022) |
| Segmentation and localization | Whole-mammogram mass segmentation, ROI segmentation, weakly supervised segmentation from classification, or mass detection | (Sun et al., 2018, Bozorgpour et al., 3 Jun 2026, Ma et al., 6 Aug 2025, Famouri et al., 2021) |
| Multimodal and systems studies | Metadata distillation, descriptor fusion, VQA, federated learning, and harmonization | (Baur et al., 6 Aug 2025, Ben-Artzi et al., 2024, Li et al., 15 Aug 2025, Zhu et al., 9 Jan 2026, Zafari et al., 4 Nov 2025) |
Whole-image and breast-level classification studies include a score-based generative classifier operating on whole mammograms with benign-versus-malignant labels (Sarker et al., 24 Feb 2025), a ViT-GNN framework for normal-versus-abnormal image classification (Cai et al., 11 Jul 2025), and a three-stage EfficientNet system that culminates in two-view same-breast classification from CC and MLO mammograms (Petrini et al., 2021). Multi-view and dual-scale classification variants also appear: Mammo-Mamba uses four images per breast examination—cropped and whole CC plus cropped and whole MLO—on a mass-only binary task (Bayatmakou et al., 23 Jul 2025), while Mammo-Clustering uses four standard views per patient with weakly supervised lesion localization integrated into case-level benign-versus-malignant prediction (Yang et al., 8 Jul 2025).
ROI-centric studies use CBIS-DDSM as a lesion-characterization benchmark rather than a screening corpus. The EfficientNetV2-M plus Vision Mamba hybrid treats the dataset as abnormality-centered cropped ROIs and classifies each ROI as benign or malignant in a patient-level stratified split (Asad et al., 14 Apr 2026). The handcrafted-feature fusion paper also uses cropped ROI images from both masses and calcifications for benign-versus-malignant classification (Tschuchnig et al., 26 Jul 2025). Calcification-specific work narrows further to cropped calcification regions for two-class and three-class management/pathology tasks (Panambur et al., 2022).
Segmentation work spans full mammograms, lesion ROIs, and weak supervision. AUNet uses the mass subset and segmentation masks to segment breast masses directly from whole mammograms (Sun et al., 2018). Swin-SFTNet builds a custom CBIS-DDSM micro-mass benchmark from whole mammograms resized to 256×256 and tests segmentation under extremely small-lesion conditions (Kamran et al., 2022). DSU-Net instead operates on cropped lesion images and binary ROI masks, mixing mass and calcification cases in a patient-wise ROI segmentation setting (Bozorgpour et al., 3 Jun 2026). ExplainSeg occupies a different regime: it trains only with image-level class labels, derives relevance maps with Integrated Gradients, and evaluates the resulting segmentation masks against held-out lesion masks (Ma et al., 6 Aug 2025). This breadth of use suggests that CBIS-DDSM is less a single benchmark than an umbrella resource supporting fundamentally different problem statements.
4. Split conventions and preprocessing regimes
No single protocol governs CBIS-DDSM usage. Some papers adopt official or recommended splits; others construct patient-wise partitions, lesion-type-specific subsets, or custom hard benchmarks. The multimodal privileged distillation study states only that it uses “official CBIS-DDSM datasplits,” with no further counts (Baur et al., 6 Aug 2025). The hybrid EfficientNetV2-M plus Vision Mamba classifier uses a patient-level stratified 70%/15%/15% train/validation/test split (Asad et al., 14 Apr 2026). The foundation-model transfer study applies a uniform 80%/10%/10% split across datasets but does not report whether this is patient-level for CBIS-DDSM (Mansoori et al., 26 May 2025). The multimodal descriptor-fusion network uses five-fold stratified cross-validation for mass lesions (Ben-Artzi et al., 2024). The two-view EfficientNet paper reports both five-fold cross-validation and the original dataset division, excluding one-view cases from the final two-view stage (Petrini et al., 2021). Other works define custom task splits such as 849 training and 69 test images for a micro-mass challenge (Kamran et al., 2022), or patient-level 80%/20% splitting for ROI segmentation (Bozorgpour et al., 3 Jun 2026).
Preprocessing conventions are likewise heterogeneous. Common operations include grayscale-to-RGB duplication for ImageNet-pretrained backbones, resizing, and train-time augmentation, but neither the spatial resolution nor the breast-region handling strategy is standardized. ROI classifiers use 224×224 (Asad et al., 14 Apr 2026), 512×512 or 600×600 after resizing and cropping (Tschuchnig et al., 26 Jul 2025), or 224×224 calcification patches (Panambur et al., 2022). Full-image or multi-view models may use 1152×896 mammograms (Petrini et al., 2021), 1024×1024 inputs (Ben-Artzi et al., 2024), or fixed 224×224 inputs for foundation models (Mansoori et al., 26 May 2025). Whole-mammogram segmentation work often resizes aggressively to 256×256 after background trimming (Sun et al., 2018, Kamran et al., 2022).
Laterality and orientation handling have emerged as a distinct preprocessing concern. Mammo-Mamba states that images are uniformly flipped to address variability between left and right breast scans (Bayatmakou et al., 23 Jul 2025). MammoClean formalizes this into a harmonization rule: right-laterality images are horizontally flipped to enforce a common orientation, and approximately 28% of CBIS-DDSM images are reported to have flipped laterality issues (Zafari et al., 4 Nov 2025). Background removal is task-dependent: Mammo-Clustering converts DICOM to 16-bit PNG, applies Canny edge detection, morphology, and contour-based cropping to isolate breast regions (Yang et al., 8 Jul 2025), whereas several other papers do not report any explicit breast masking or pectoral removal. The literature therefore presents preprocessing not as a stable CBIS-DDSM standard but as a major source of inter-study variability.
5. Benchmark behavior and empirical patterns
CBIS-DDSM has produced both strong headline results and intentionally negative findings, depending on the formulation. On lesion-centered ROI classification, the EfficientNetV2-M plus Vision Mamba hybrid reports test-set performance of AUC 0.875, accuracy 94.2%, sensitivity 0.89, specificity 0.95, and F1 0.90, outperforming the evaluated baselines in that ROI-based protocol (Asad et al., 14 Apr 2026). In transfer-learning experiments with natural-domain foundation models, the best reported CBIS-DDSM ROC-AUC is 0.968 for AIMv2 ViT-L with an unfrozen linear head, with DINOv2 models close behind at 0.966, although the exact experimental label formulation is not fully specified (Mansoori et al., 26 May 2025). Mammo-Mamba reports 0.9089 ± 0.0135 AUC, 0.8696 ± 0.0109 accuracy, and 0.8396 ± 0.0124 F1 on its curated mass-only dual-view, dual-scale classification subset (Bayatmakou et al., 23 Jul 2025).
Segmentation results also vary sharply with task difficulty. AUNet reports an average Dice similarity coefficient of 81.8% for CBIS-DDSM whole-mammogram breast mass segmentation, exceeding the compared FCN baselines under its 690/168 train/validation protocol (Sun et al., 2018). By contrast, the custom micro-mass benchmark in Swin-SFTNet yields only 24.13% Dice and 17.44% mIoU for the proposed method, despite outperforming AUNet and SWIN-UNet on that harder tiny-lesion setting (Kamran et al., 2022). ExplainSeg, which derives masks from classification explanations without mask supervision during training, reaches 31.2% mIoU and 43.7% Dice with its XAI+NCut variant, substantially outperforming TokenCut, MICRA-Net, and MaskCut on CBIS-DDSM while still remaining far below conventional supervised-segmentation performance (Ma et al., 6 Aug 2025). This contrast indicates that “performance on CBIS-DDSM” is only meaningful when the supervision regime and lesion scale are specified.
CBIS-DDSM has also been central to papers whose main contribution is a negative or cautionary result. In multimodal privileged knowledge distillation, metadata-only teachers achieve AUROC 0.86 for mass and 0.90 for calcification, yet the ViT-Tiny student shows only stable predictive AUROC and degraded attention-based zero-shot ROI localization under distillation, leading the authors to argue that privileged distillation effects do not generalize across domains (Baur et al., 6 Aug 2025). Federated-learning experiments similarly treat CBIS-DDSM as a harder medical transfer case: RHFL+ is reported as inconsistent on binary benign/malignant and four-class density tasks, with lower-than-expected overall accuracy and only partial recovery after switching to an EfficientNet-based model (Zhu et al., 9 Jan 2026). A plausible implication is that CBIS-DDSM is especially useful when a method’s robustness claims need to be stress-tested against dataset ambiguity, small-sample regimes, and multi-view inconsistency.
6. Bias, ambiguity, and reproducibility issues
CBIS-DDSM’s utility is inseparable from its reproducibility challenges. Many papers omit crucial details: exact retained sample counts after filtering, patient-level versus image-level partitioning, view-completeness criteria, loss instantiations, or metadata fields actually consumed. The multimodal privileged distillation study does not list the metadata columns used by its random-forest teacher and does not specify partition granularity beyond “official datasplits” (Baur et al., 6 Aug 2025). The GPT-5 mammogram VQA benchmark uses balanced sampling over automatically generated question types and answer categories but does not disclose the exact CBIS-DDSM evaluation subset, split protocol, or preprocessing pipeline (Li et al., 15 Aug 2025). DSU-Net gives a patient-wise 80%/20% split and a detailed training recipe, yet does not disclose the final number of valid image-mask pairs retained from the training CSVs (Bozorgpour et al., 3 Jun 2026). Such omissions make cross-paper comparison difficult even when the dataset name is shared.
The most systematic critique comes from MammoClean. That paper reports patient overlap between the predefined CBIS-DDSM training and test sets, leading the authors to discard the predefined split for their unified storage scheme (Zafari et al., 4 Nov 2025). It also identifies approximately 28% of images with flipped laterality issues and shows that CBIS-DDSM, after harmonization, has no normal cases in its diagnosis distribution and a BI-RADS distribution dominated by category 4 (Zafari et al., 4 Nov 2025). The same study highlights modality mismatch—CBIS-DDSM is screen-film mammography, whereas some contemporary public datasets are full-field digital mammography—and notes that age is not available in CBIS-DDSM’s harmonized metadata. These observations matter because they imply that models trained on CBIS-DDSM may inherit suspicious-case enrichment, density-distribution shift, and orientation artifacts unless such issues are explicitly corrected.
Another recurrent limitation is that the dataset’s semantic richness can become a source of task-definition drift. Papers variously treat it as a mass-only set, a mass-and-calcification ROI corpus, a whole-image benign/malignant dataset, a normal/abnormal image-level benchmark, or a four-view patient-level resource (Famouri et al., 2021, Tschuchnig et al., 26 Jul 2025, Sarker et al., 24 Feb 2025, Cai et al., 11 Jul 2025, Yang et al., 8 Jul 2025). This suggests that reported CBIS-DDSM performance figures are only comparable after fixing the lesion subset, unit of prediction, view aggregation, split policy, and use of ROI or mask information. For encyclopedia purposes, the most accurate characterization is therefore that CBIS-DDSM is a curated but still structurally heterogeneous mammography benchmark whose scientific value lies as much in exposing protocol sensitivity as in providing a common public dataset.