BscanFound

BscanFound is a domain-specific visual foundation model pretrained on ocular B-scan ultrasound data using Masked Autoencoding (MAE).
The released checkpoint uses a ViT-Base/16 encoder and is intended to provide transferable representations for downstream ocular-ultrasound tasks.

Important: This model is a research representation model. It is not a standalone clinical diagnostic system and has not been validated for autonomous clinical use.

Model Summary

Item	Value
Architecture	ViT-Base/16 (vit_base_patch16)
Pretraining objective	Masked Autoencoding (MAE)
Mask ratio	75%
Pretraining epochs	400
Effective batch size	256
Training precision	BF16
Training hardware	4 × NVIDIA GeForce RTX 4090
Total training time	5.96 hours
Best validation loss	0.400514
Final validation loss	0.401418

The released checkpoint_best.pth corresponds to the checkpoint with the lowest validation reconstruction loss during pretraining.

Pretraining Data

BscanFound was pretrained using two ocular B-scan ultrasound datasets: ERDES and Oculo.

ERDES

ERDES is a video-based ocular ultrasound dataset. To reduce temporal redundancy, videos were sparsely sampled before MAE pretraining.

• Sampling rate: 2 FPS
• Maximum sampled frames per video: 12
• Official ERDES train/validation partitions were retained.
• The official ERDES test set was excluded from MAE pretraining to preserve it for downstream evaluation.

ERDES split	Samples used for MAE
Train	26,057 frames
Validation	2,841 frames

Oculo

Oculo contains static ocular B-scan images. Diagnostic labels were not used during MAE pretraining.

The images were randomly divided into an 80% training / 20% validation split.

Oculo split	Samples used for MAE
Train	1,304 images
Validation	326 images

Combined Pretraining Corpus

Split	Total samples
Train	27,361
Validation	3,167

ERDES contributes approximately 95.2% of the training samples, so the current representation is predominantly shaped by the ERDES ultrasound domain.

Training

The model was pretrained from ocular B-scan images using the standard MAE formulation:

1. Each input image is divided into non-overlapping ViT patches.
2. 75% of patches are randomly masked.
3. The ViT encoder processes only the visible patches.
4. A lightweight MAE decoder reconstructs the masked image patches.
5. Reconstruction loss is optimized only over masked patches.

Training completed all 400 epochs without instability.

Training Loss

The training reconstruction loss decreased continuously throughout pretraining, with a slower but persistent improvement during the later epochs.

Validation Loss

Validation loss decreased rapidly during early training and gradually approached a plateau near 0.40. The best validation loss was 0.400514, while the final validation loss was 0.401418, only about 0.23% higher than the best value.

This behavior suggests that the MAE reconstruction objective had largely converged by the end of the 400-epoch schedule, with only mild late-stage saturation.

Intended Use

BscanFound is intended primarily as a pretrained visual backbone for research on ocular B-scan ultrasound.

Potential downstream applications include:

• ocular B-scan image classification;
• retinal detachment recognition;
• posterior vitreous detachment recognition;
• macular status classification;
• multi-label ultrasound finding recognition;
• feature extraction and linear probing;
• fine-tuning for other ocular ultrasound datasets;
• initialization for multimodal or vision-language models using ocular B-scan images.

The checkpoint is expected to be most useful when fine-tuned or probed on task-specific labeled data.

Limitations

Several limitations should be considered when using this checkpoint.

• Dataset imbalance: ERDES accounts for approximately 95.2% of the training samples.
• Temporal correlation: Multiple training images can originate from the same ultrasound video and are therefore not statistically independent.
• Limited domain diversity: The current pretraining corpus contains two ocular B-scan datasets and should not be assumed to represent all devices, institutions, acquisition protocols, or patient populations.
• MAE loss is not a clinical metric: Lower reconstruction loss does not directly imply better diagnostic accuracy. Downstream evaluation is required.
• No autonomous clinical validation: The model should not be used as a standalone diagnostic system.

Checkpoint

The main released model file is:

checkpoint_best.pth

This checkpoint was selected according to the lowest validation MAE reconstruction loss observed during the 400-epoch pretraining run.

Because the checkpoint follows the custom BscanFound MAE implementation, users should instantiate the corresponding vit_base_patch16 MAE architecture before loading the weights.

Recommended Evaluation

For downstream experiments, we recommend comparing BscanFound against at least:

1. the same ViT-B/16 architecture initialized randomly;
2. a general-domain ImageNet-pretrained ViT-B/16;
3. BscanFound MAE-pretrained ViT-B/16.

This comparison can determine whether ocular B-scan-specific self-supervised pretraining provides transferable benefit beyond generic visual pretraining.

Training Run Summary

{
  "model": "vit_base_patch16",
  "epochs_completed": 400,
  "train_samples": 27361,
  "val_samples": 3167,
  "train_source_counts": {
    "ERDES": 26057,
    "Oculo": 1304
  },
  "effective_batch_size": 256,
  "best_val_loss": 0.40051395024615105,
  "last_val_loss": 0.40141790307531455,
  "amp_dtype": "bf16",
  "total_hours": 5.96448813021183
}

Citation

A formal citation for BscanFound will be added when the corresponding work is released.