TesticulUS: Testicular B-mode Ultrasound Dataset

@article{,
title= {TesticulUS: Testicular B-mode Ultrasound Dataset},
keywords= {Image classification, ultrasound, Image Segmentation, Medical imaging, testicular ultrasound, synthetic images, diffusion model},
author= {},
abstract= {# TesticulUS

A public collection of testicular ultrasound images for medical image analysis, from the AImageLab group at the University of Modena and Reggio Emilia (UNIMORE). It has two parts: a synthetic set for classification pretraining, and a real multicenter set for segmentation.

## TesticulUS-Syn (ICIAP 2025)

Synthetic B-mode images generated with a Denoising Diffusion Probabilistic Model (DDPM). They were released because the real clinical images behind the original classification study could not be shared. The study targeted classification of parenchymal inhomogeneity and proposed a training pipeline built around synthetic pretraining.

| Field | Value |
| --- | --- |
| Organ | Testicle |
| Images | 9,289 |
| Generation method | DDPM, custom implementation based on OpenAI guided-diffusion |
| File format | PNG |
| Image shape | 256 x 256 |
| Modality | Ultrasound, B-mode (synthetic) |
| License | CC BY-NC-SA |

The images were validated and filtered using the technique described in the paper.

## TesticulUS-Real (BMVC 2026)

The first multicenter testicular ultrasound segmentation dataset. Each image comes with an expert-reviewed segmentation mask. A subset also carries the parenchymal inhomogeneity classification label from the ICIAP 2025 study, so it supports both segmentation and classification work.

| Field | Value |
| --- | --- |
| Organ | Testicle |
| Images | 1,053 |
| Centers | 2 (Baggiovara Hospital, Modena; Umberto I Hospital, Rome) |
| Acquisition systems | Esaote MyLab25 Gold / MyLab XPro80 (851 images), Philips EPIQ 5 (202 images) |
| File format | PNG |
| Modality | Ultrasound, B-mode |
| License | CC BY-NC-SA |

The accompanying paper also introduces Cond-UNet, a lightweight conditional U-Net using FiLM modulation and a Shared Attention Conditioning (SAC) mechanism. It was benchmarked across five anatomical regions with the testicular data held out as an unseen organ, and it matches or beats much larger foundation-model baselines at a fraction of the cost.

## Resources

- Cond-UNet code: https://github.com/AImageLab-zip/US_Cond-UNet
- Classification code: https://github.com/AImageLab-zip/TesticulUS
- Cond-UNet on Hugging Face: https://huggingface.co/AImageLab-Zip/US_Cond-UNet

## Citation

Anyone using the dataset must cite both papers:

- Enhancing Testicular Ultrasound Image Classification through Synthetic Data and Pretraining Strategies (ICIAP 2025)
- A New Multicenter Testicular US Dataset and a Lightweight Cond-UNet for Generalization in US Segmentation (BMVC 2026)

},
terms= {},
license= {https://creativecommons.org/licenses/by-nc-sa/4.0/},
superseded= {},
url= {https://ditto.ing.unimore.it/testiculus/}
}