# Evidence and technical choices

The project concerns benign/malignant classification and index-lesion segmentation in ultrasound. This is a focused review of primary dataset papers, architecture papers and reporting guidance, not a systematic review or a claim of state-of-the-art performance. Published scores are not treated as comparable when patient units, splits, ROI inputs or labels differ.

## Choice of modality and data

Ultrasound provides diagnosis labels and lesion contours within a modest download. BUS-BRA contains 1,875 images from 1,064 released patient cases [1,2]. BrEaST contains 256 patient images, including 252 index lesions and four normal cases [3,4]. These data permit both tasks with patient-grouped development and native-mask scoring. The adapter uses pathology rather than BI-RADS as the malignant/benign target.

| Modality | Relevant public resources | Fit to this project |
| --- | --- | --- |
| Ultrasound | BUS-BRA; BrEaST | Diagnoses, index masks and released patient groups; manageable 2D inputs. Speckle, acquisition differences and on-image markers complicate inference. |
| Mammography | CBIS-DDSM; VinDr-Mammo | Valuable for lesion analysis, but larger collections and multiple views require a different scope. BI-RADS alone is not a histological malignancy label. |
| Histopathology | BreakHis; BACH | Tissue classification is a different image scale and workflow. BreakHis has 82 patients and no corresponding tumor masks for this segmentation task. |

Official source links: [CBIS-DDSM](https://www.cancerimagingarchive.net/collection/cbis-ddsm/), [VinDr-Mammo](https://physionet.org/content/vindr-mammo/1.0.0/), [BreakHis descriptor](https://www.inf.ufpr.br/lesoliveira/download/TBME-00608-2015-R2-preprint.pdf), [BACH paper](https://arxiv.org/abs/1808.04277). Ultrasound is selected for its paired diagnosis/mask annotations and feasible resource requirements, rather than because results across these modalities establish an intrinsic advantage.

BUS-BRA `Case` is the grouping unit; `ID` denotes images. The original classification code uses reference-mask-assisted input, which differs from complete-image inference here. Released outer folds keep patients together, but validation columns share patient views with training. Our custom split and audit address that issue [DATA.md](data.md).

BrEaST has biopsy confirmation for malignant cases and some benign cases; other benign cases use follow-up. It annotates the index lesion separately from other findings. Those extra masks are retained for error analysis without assigning the case diagnosis to every object. Equipment, acquisition and annotation differences can affect comparisons between the collections.

## Model family

ResNet [5] supplies a compact, well-defined pretrained encoder. Comparing frozen features with supervised fine-tuning tests whether task-specific optimization adds useful information. ResNet18 classification is economical for the available sample size. ResNet18 and ResNet34 segmentation encoders provide a finite comparison of capacity within the same decoder design. ImageNet weights are a starting point, not prior medical validation.

U-Net [6] combines a contracting encoder with spatial decoding and skip connections. This implementation retains that design principle with a residual encoder and five decoder stages; it is not an exact reproduction of the original microscopy network. The decoder, input sizes, optimizer and augmentation settings are engineering choices documented in [METHODS.md](methods.md).

Transformers and large ultrasound foundation models are relevant context, but choosing them would require a justified pretrained source, additional compute and an evaluation that distinguishes initialization from architecture. For example, BUSGen uses ultrasound-scale pretraining far beyond these two public cohorts [9]. Such a model is not an appropriate reference for claiming that a small, ImageNet-initialized experiment reaches the best published performance. The present benchmark makes a limited comparison of reproducible baselines and models it actually trains.

## Reporting and comparison

CLAIM 2024 [7] supports explicit description of data, partitions, acquisition, model development and evaluation. Metrics Reloaded [8] emphasizes selecting metrics and aggregation units that match the task. Accordingly, classification uses case-level ranking, decisions and calibration; segmentation uses native-image overlap aggregated by case. Lower-tail scores and error examples complement averages.

A selected lesion collection cannot establish screening sensitivity or specificity for absence of disease. A mask-derived box is localization of the annotated index lesion, not exhaustive object detection. One patient split and three seeds quantify only part of experimental variation. Scores from the other collection describe transfer under its released case composition; broader conclusions need further data.

## Primary references

1. Gómez-Flores W, Gregorio-Calas MJ, Pereira WCA. *BUS-BRA: A breast ultrasound dataset for assessing computer-aided diagnosis systems*. Medical Physics, 2024. [DOI 10.1002/mp.16812](https://doi.org/10.1002/mp.16812).
2. [Official BUS-BRA data, Zenodo 8231412](https://zenodo.org/records/8231412); [authors' source repository](https://github.com/wgomezf/BUS-BRA).
3. Pawłowska A et al. *Curated benchmark dataset for ultrasound based breast lesion analysis*. Scientific Data 11, 148, 2024. [DOI 10.1038/s41597-024-02984-z](https://doi.org/10.1038/s41597-024-02984-z).
4. [BrEaST / Breast-Lesions-USG, TCIA](https://www.cancerimagingarchive.net/collection/breast-lesions-usg/); [official importer](https://github.com/best-ippt-pan-pl/BrEaST).
5. He K et al. *Deep Residual Learning for Image Recognition*. CVPR, 2016. [Paper](https://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html). [TorchVision ResNet34 weights](https://docs.pytorch.org/vision/stable/models/generated/torchvision.models.resnet34.html); implemented weights are explicitly pinned to `IMAGENET1K_V1` in TorchVision 0.25.0.
6. Ronneberger O, Fischer P, Brox T. *U-Net: Convolutional Networks for Biomedical Image Segmentation*. MICCAI, 2015. [Paper](https://arxiv.org/abs/1505.04597).
7. Tejani AS et al. *Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 update*. Radiology: Artificial Intelligence. [DOI 10.1148/ryai.240300](https://doi.org/10.1148/ryai.240300).
8. Maier-Hein L et al. *Metrics Reloaded: recommendations for image analysis validation*. Nature Methods, 2024. [DOI 10.1038/s41592-023-02151-z](https://doi.org/10.1038/s41592-023-02151-z).
9. *BUSGen*, Nature Biomedical Engineering, 2026. [Primary paper](https://www.nature.com/articles/s41551-026-01639-1).

Original notices: [BUS-BRA](BUSBRA_LICENSE.txt) and [BrEaST](BREAST_LICENSE.txt). Data checksums, verification details and annotation contracts are recorded in [DATA.md](data.md).