mBRSET: Mobile Brazilian Retinal Dataset

• mBRSET is a smartphone-based retinal dataset comprising 5,164 images from 1,291 patients, specifically designed to evaluate device variability in diabetic retinopathy (DR) and diabetic macular edema (DME) diagnosis.
• The dataset complements lab-captured BRSET by sharing key clinical metadata, enabling consistent multimodal fusion and demonstrating approximately 4% improvement in balanced accuracy with MetaFusion pretraining.
• mBRSET supports a lightweight mobile screening pipeline using a ResNet-18 backbone and a real-time quality checker to filter low-quality images, ensuring robust performance across varying imaging conditions.

Searching arXiv for the cited paper and any related mBRSET references. mBRSET, the Mobile Brazilian Retinal Dataset, is a dataset of smartphone-based retinal fundus camera images used to evaluate robustness on smartphone-captured fundus images and to assess generalization across different imaging conditions in the InSight mobile screening pipeline. Within that study, mBRSET complements BRSET, which contains lab-captured fundus images, while sharing key clinical metadata so that multimodal fusion can be trained consistently across domains. The dataset is Brazilian, publicly available via PhysioNet, and is used in InSight primarily for diabetic retinopathy (DR) and diabetic macular edema (DME) prediction rather than for the full five-disease label space available in BRSET (Raghu et al., 16 Jul 2025).

1. Dataset identity and role in multimodal screening

mBRSET is positioned as the mobile-capture counterpart to BRSET. Its stated purpose is to evaluate InSight’s robustness on smartphone-captured fundus images and to assess generalization across different imaging conditions. In the reported workflow, it represents the mobile capture domain, whereas BRSET represents the lab-captured domain. The study limited metadata usage to fields present in both datasets so that a unified multimodal model could be trained across the two domains (Raghu et al., 16 Jul 2025).

This pairing is central to the design of InSight. The pipeline is described as a three-stage system comprising real-time image quality assessment, a disease diagnosis model, and a DR grading model to assess severity. mBRSET enters that framework as the smartphone-captured dataset used for disease diagnosis experiments on DR and DME and for evaluating cross-device robustness. A common misunderstanding is to treat mBRSET as a five-disease benchmark analogous to BRSET. The paper does not support that interpretation: BRSET includes age-related macular degeneration, glaucoma, DR, DME, and pathological myopia, whereas mBRSET includes DR and DME.

A plausible implication is that mBRSET’s principal value in this study is not disease-label breadth but domain diversity. The dataset is therefore used to test how multimodal fusion, pretraining, and quality filtering behave when image acquisition shifts from lab-captured to smartphone-captured fundus images.

2. Composition, labels, and ground-truth scope

mBRSET contains 5,164 images from 1,291 patients. The paper states that it contains labels for diabetic retinopathy (DR) and diabetic macular edema (DME) with additional clinical information (Raghu et al., 16 Jul 2025). It does not report per-disease counts for mBRSET or detailed prevalence rates.

The manuscript also does not state whether mBRSET uses strictly single-label per image or multi-label annotation. However, DR and DME are treated as separate tasks. This suggests potential multi-label assignments, but that inference should not be read as an explicit annotation specification.

The scope of available ground-truth documentation in the paper is limited. The study does not describe the grading protocol for mBRSET, the number of graders, inter-rater agreement, or gold standards. It states that mBRSET provides labels for DR and DME but omits annotation methodology details. Likewise, the paper does not report DR severity grading results on mBRSET. The DR severity grading model, described as covering mild nonproliferative through severe proliferative disease, was trained and evaluated on BRSET only.

These omissions delimit what can be claimed about mBRSET from this source. The dataset is clearly established as a smartphone fundus dataset with DR and DME labels, but annotation provenance and detailed prevalence structure are not documented in the manuscript.

3. Shared metadata and preprocessing logic

The model uses the subset of metadata shared between BRSET and mBRSET. The fields used in the study are:

Field	Representation
Age	years
Sex	binary
Diabetes diagnosis	binary
Duration of diabetes	years
Hypertension	binary; extracted from comorbidities text

The use of this shared subset is methodologically important because it enables unified multimodal fusion across the lab-captured and smartphone-captured domains (Raghu et al., 16 Jul 2025).

The paper reports several preprocessing details, though not all of them are mBRSET-specific. In BRSET, one-third of patients had missing age; this was imputed using the median age. The manuscript does not report missingness for age or other fields in mBRSET. Hypertension was not reported as a standalone field in BRSET and was instead extracted as a binary feature from comorbidities text. The text implies consistent preprocessing across the shared metadata fields used by both datasets.

The study also applied consistency corrections where DR or DME labels conflicted with diabetes status, for example by enforcing that DR implies diabetes, and DME implies both diabetes and DR. The manuscript describes these corrections broadly and does not separate whether they were needed in mBRSET specifically, but the intent was to enforce logical consistency across the used datasets.

This metadata design has direct modeling consequences. InSight’s multimodal fusion mechanism depends on overlap in metadata schema rather than on dataset-specific side information, which constrains the feature set but supports cross-domain training.

4. Imaging modality and image-quality constraints

mBRSET consists of smartphone-based retinal fundus camera images. The paper describes InSight as a mobile app pipeline and cites prior smartphone imaging systems, but for mBRSET specifically it only states “smartphone-based retinal fundus cameras” and does not name a particular make, model, optics, or capture app (Raghu et al., 16 Jul 2025). Specific clinic or community settings, operator characteristics, and capture protocols are likewise not described.

Several image-level characteristics are explicitly not reported for mBRSET in the paper: resolution, format, color space, field of view, and eye laterality markers. What the manuscript does emphasize is smartphone-specific quality variability, particularly blurriness and altered brightness. These conditions were explicitly simulated through augmentations—MotionBlur, GaussianBlur, MedianBlur, RandomBrightnessContrast—when training the real-time image quality checker.

The quality checker is a binary CNN classifier trained using augmented low-quality images and original images in the joint training setting with BRSET and mBRSET. The reported result is near-100% accuracy in filtering out low-quality images across both BRSET and mBRSET. Thresholds and quantitative precision/recall are not reported. The operational criterion is that the checker rejects images deemed blurry or with altered brightness; images that pass the filter proceed to diagnosis, while DR grading is only reported on BRSET.

A plausible implication is that mBRSET’s contribution is not merely as an additional dataset but as a source of mobile-image degradation modes that directly motivate the quality-control stage.

5. Experimental protocols and benchmarked performance

The paper reports several training and evaluation configurations involving mBRSET. BRSET was split 0.5/0.25/0.25 (train/val/test), and for the combined dataset (BRSET + mBRSET) the same ratio was adopted. The manuscript does not state subject-level separation or fold composition for mBRSET alone. Evaluation within the paper is internal to BRSET and mBRSET; no external datasets were used for smartphone testing beyond mBRSET (Raghu et al., 16 Jul 2025).

Three training regimes are described: models trained on BRSET alone, on mBRSET alone, and on the joint dataset. The joint model performed “almost as well” as device-specific models, which the paper interprets as evidence of cross-domain robustness.

For mBRSET, the reported disease diagnosis results are balanced accuracies (BA):

Training setup	DR BA	DME BA
Trained on mBRSET, tested on mBRSET	0.79	0.87
Trained on BRSET + mBRSET, tested on mBRSET	0.79	0.86

The paper also states that MetaFusion + pretraining improves BA on mBRSET by approximately +4% for DR and for DME relative to image-only baselines, and summarizes this as “improving accuracy by 4% for DR and ME.” By contrast, AUC, F1, specificity, and sensitivity on mBRSET are not reported. DR severity grading on mBRSET is also not evaluated or reported.

The balanced accuracy formulas given in the study are:

$BA = \frac{TPR + TNR}{2}$

for the binary case, and

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$

for the multiclass case.

These results situate mBRSET as a domain-robustness benchmark rather than as a comprehensive ophthalmic benchmark. The reported numbers are specifically tied to DR and DME classification.

6. Model architecture, fusion mechanism, and deployment relevance

The methods tied to mBRSET in InSight include multimodal fusion, pretraining, multitask learning, and mobile-oriented efficiency choices. The disease diagnosis model uses ResNet-18 as the image backbone. The paper states that one embedding is created per metadata feature, aligned to the image embedding dimension $d$, and then fused through a residual correction mechanism (Raghu et al., 16 Jul 2025).

Using the notation given in the paper, where “@” denotes element-wise product, the correction term for embeddings $a$ and $b$ of dimension $d$ is

$f(a, b) = a @ \tanh(a @ W b)$

where $W \in \mathbb{R}^{d \times d}$ maps $b$ into the dimension of $a$.

The image embedding is then corrected using all $$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$0 metadata embeddings:

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$1

Each metadata embedding is corrected using the image:

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$2

The corrected embeddings are concatenated and fed to task-specific linear layers.

For the multitask setting across five diseases on BRSET, with mBRSET contributing to DR and DME tasks, the loss is

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$3

The pretraining objective combines supervised classification for DR and glaucoma with self-supervised image reconstruction:

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$4

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$5

$$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$6

The paper does not specify $$BA = \frac{1}{C} \sum_{c=1}^{C} \frac{TP_c}{TP_c + FN_c}$$7.

From a deployment standpoint, the multitask diagnosis model—one ResNet-18 backbone with five task heads—is described as approximately 5× more computationally efficient than training and deploying five separate single-disease models, while achieving comparable performance. ResNet-18 was chosen for lightweight inference on smartphones. The quality checker prevents low-quality inputs from propagating, and the study states that multimodal fusion stabilizes predictions by leveraging metadata when images are suboptimal. This suggests that mBRSET is integral to validating both algorithmic and systems-level claims about mobile screening.

7. Access, citation, and documented limitations

mBRSET is publicly available via PhysioNet at:

https://physionet.org/content/mbrset/1.0/

The dataset is cited in the paper as:

Nakayama LF, Santos F, Barbosa I, Pereira R, Lima R, Oliveira C, et al. mBRSET, a mobile Brazilian retinal dataset [dataset]. PhysioNet. 2024. Available from: https://physionet.org/content/mbrset/1.0/ (Raghu et al., 16 Jul 2025)

The InSight code is available at:

https://github.com/Anisha234/InSight

The paper does not specify licensing terms, access approvals, or data use agreements for mBRSET; consultation of the PhysioNet page is therefore necessary for current data use policies.

Several limitations are explicit. The manuscript does not report mBRSET per-class prevalence, image resolution, format, color space, field of view, eye laterality markers, annotation protocols, number or type of graders, inter-rater statistics, or AUC/F1/sensitivity/specificity on mBRSET. It also does not specify the smartphone hardware beyond the phrase “smartphone-based retinal fundus cameras.” A common misconception is that these details can be recovered from the InSight paper alone. They cannot; the manuscript explicitly leaves them unspecified.

Taken together, these characteristics define mBRSET as a public smartphone fundus dataset used in multimodal, cross-device ophthalmic screening research. Its principal documented function in the cited work is to test whether mobile fundus images, paired with a shared subset of clinical metadata, can support robust DR and DME prediction within a lightweight screening pipeline.