TesticulUS

Overview

TesticulUS is a public dataset for testicular ultrasound image analysis, built around two complementary studies from our group. It was originally introduced in Enhancing Testicular Ultrasound Image Classification through Synthetic Data and Pretraining Strategies (ICIAP 2025), which targeted the classification of parenchymal inhomogeneity but could not release the underlying clinical images; instead, a synthetic dataset generated with a Denoising Diffusion Probabilistic Model (DDPM) was released to support pretraining and further research.

With A New Multicenter Testicular US Dataset and a Lightweight Cond-UNet for Generalization in US Segmentation (BMVC 2026), we are now able to publicly release the real clinical data: TesticulUS-Real, the first multicenter testicular ultrasound segmentation dataset. It contains 1,053 real B-mode ultrasound images collected from two independent clinical institutions, each paired with an expert-reviewed segmentation mask. A subset of these images also carries the parenchymal-inhomogeneity classification label defined in the ICIAP 2025 study, so the same release now supports both segmentation and classification research.

TesticulUS-Syn - ICIAP2025

In our ICIAP 2025 paper, we investigate the automated classification of testicular parenchymal inhomogeneity in ultrasound images. We examine the performance achievable with current classification models and discuss the key challenges encountered when working with limited clinical data. We also introduce a training pipeline that achieved our best classification results. To support further research in this underexplored field, we trained a denoising diffusion probabilistic model (DDPM) to generate synthetic testicular ultrasound images and evaluated their usefulness within our classification pipeline.

The proposed training pipeline.

Field	Value
Organ	Testicle
# images	9,289
Generation method	Denoising Diffusion Probabilistic Model (DDPM), custom implementation based on OpenAI guided-diffusion
File format	PNG
Image shape	256 × 256
Modality	Ultrasound / B-mode (synthetic)
License	CC BY-NC-SA

 

TesticulUS-Real - BMVC2026

Public datasets for testicular ultrasound segmentation are essentially non-existent, and models trained at a single hospital often struggle to generalize to images from other scanners and centers. This BMVC 2026 paper introduces TesticulUS-Real, the first multicenter testicular ultrasound segmentation dataset, together with Cond-UNet, a lightweight conditional U-Net that combines FiLM modulation with a novel Shared Attention Conditioning (SAC) mechanism to adapt its representations to the organ being segmented.

Benchmarked across five anatomical regions — with the testicular data held out as a completely unseen organ — Cond-UNet matches or outperforms much larger foundation-model baselines while using a fraction of their parameters and computational cost. The real clinical images and expert-reviewed segmentation masks collected for this study are released as the TesticulUS-Real dataset, presented below.

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

 

Download

You need to have an account to download the dataset. Please sign in or sign up!

How to cite

To cite the TesticulUS dataset, please use the following references:

If you use our dataset, you must cite the following papers.