BUS-BRA: A Breast Ultrasound Dataset for Assessing Computer-Aided Diagnosis Systems

Paper. BUS-BRA: A breast ultrasound dataset for assessing computer-aided diagnosis systems.

BUS-BRA addresses a problem that architecture papers cannot solve on their own: how can breast-ultrasound methods be compared when datasets differ in annotation, patient overlap, and evaluation partitions? An apparent improvement can come from a better model, but it can also come from an easier split or information that would be unavailable in the intended workflow. A reusable benchmark needs to make those choices visible.

The dataset is useful because it joins several kinds of supervision that answer different questions. Pathology supports a benign-versus-malignant target. BI-RADS records an imaging assessment. Lesion contours support spatial evaluation. Keeping these together permits investigations of relationships between diagnosis, appearance, and segmentation, while also making it necessary to distinguish their reference processes.

BUS-BRA contains 1,875 images from 1,064 women, acquired using four ultrasound scanners at the National Institute of Cancer in Rio de Janeiro. The reported patient cases comprise 722 benign and 342 malignant diagnoses. The dataset includes biopsy-proven tumors, BI-RADS categories 2 through 5, and manual lesion outlines. A senior ultrasonographer supplied the imaging assessments and contours. Standardized five- and ten-fold partitions accompany the data.

The authors also provide segmentation and classification experiments as demonstrations of use. The accompanying code includes ResNet-based classification and DeepLabV3+ segmentation frameworks. These baselines help make the benchmark executable, but the principal contribution is the documented data resource and evaluation structure. A baseline score is not the main reason to use this paper.

The distinction between images and patients is essential. There are more images than women, so observations cannot all be treated as independent new patients. A random image split could place related views on both sides of an evaluation boundary. Patient-wise partitions address that particular problem. They do not, however, prevent every other form of leakage, such as selecting preprocessing, hyperparameters, or models using the nominal test results.

Fixed folds improve comparability only when their role is respected. If a researcher repeatedly chooses methods according to performance on the same published folds, those results increasingly participate in development. Reporting the folds does not undo that information access. I would specify which data were used for model selection and reserve an independent evaluation for claims that go beyond benchmark development.

Pathology and BI-RADS also need different interpretations. A model that predicts pathology from an image is solving a different task from one that reproduces a reader’s BI-RADS assessment. Adding BI-RADS as an input changes the information available to the model and may change the intended use. It would be inappropriate to compare such a model with an image-only system as if the inputs were equivalent. Conversely, predicting BI-RADS as an auxiliary output does not prove that the pathology prediction actually uses the intended imaging findings.

The masks create another consequential design choice. A classifier given a crop derived from the reference contour is evaluated with lesion localization already supplied. That can be a legitimate study of classification conditional on a known lesion. It is not the same as evaluating a complete system that must locate the lesion in a routine examination. To understand the contribution of a proposed model, the report needs to identify whether it receives the full image, a predicted crop, or a crop based on a reference annotation.

The strongest limitation is the scope of the clinical population. Four scanners provide acquisition diversity within the collection, but they do not turn one clinical source into independent multi-institution validation. Biopsy-confirmed lesions also represent a selected clinical pathway. Performance within that pathway cannot establish screening performance among all women undergoing ultrasound, including those without a selected lesion or without biopsy verification.

There is a related annotation limitation. A contour is a reference judgment about an image boundary, while pathology establishes a diagnosis at a different level. Confidence in the pathological diagnosis does not establish certainty about every boundary pixel. The limited reader setting also does not characterize the full range of interobserver variation. These issues matter when small differences in segmentation overlap are presented as clinically meaningful gains.

BUS-BRA directly supports Data Leakage and Validation Design. Its image-to-patient structure makes the note’s distinction between prediction units and independent evaluation units concrete. The dataset also illustrates the note’s warning that a correct split is only one part of a valid evaluation procedure. Learned preprocessing and repeated selection still belong inside the development boundary.

Dataset Design, Ground Truth, and Reference Standards explains why the three annotation types should remain distinct. Selection Bias and Dataset Bias complicates the otherwise reassuring phrase “biopsy-proven”: strong verification of included lesions does not establish representativeness of patients excluded from that verification pathway.

For my work, this dataset suggests a practical sequence. I would first reproduce an appropriate published baseline using the specified patient partitions. I would then state exactly what new information or modeling component is introduced and keep the comparison inputs consistent. Results would include patient-level uncertainty and, where supported by the available metadata and counts, scanner-stratified analysis. An independent institution would be needed for a claim about transfer to that institution.

The dataset also provides a bridge to the wavelet contour paper in this corpus. Reference masks can help determine whether a more complex model improves on a transparent contour-based baseline, but the comparison should distinguish manual localization from predicted localization. BUS-BRA makes several controlled experiments possible. Its value lies in giving those experiments a clearer common basis, while leaving the researcher responsible for the clinical scope of the resulting claim.