Cond-UNet: Cross-Domain Ultrasound Segmentation

This repository contains the official PyTorch implementation for the BMVC 2026 paper: A New Multicenter Testicular US Dataset and a Lightweight Cond-UNet for Generalization in US Segmentation.

Here you will find the code to run our model in inference and to fine-tune it on your own ultrasound image dataset.

Before running the code, copy utils/paths.example.py to utils/paths.py and set the local dataset and checkpoint paths.

Overview of the proposed Cond-UNet architecture.

Dataset

TesticulUS is a multi-institutional testicular ultrasound dataset collected at two Italian clinical centers, capturing variability in acquisition protocols and devices for cross-domain evaluation.

Each image is paired with a pixel-wise expert-annotated segmentation mask. The first 860 images additionally provide labels for parenchymal inhomogeneity classification. The dataset supports:

• Cross-domain generalization experiments
• Fine-tuning and transfer learning
• Robustness benchmarking

Summary

• Images: 1,054
• Institutions: 2 Italian clinical centers
• Modality: Testicular ultrasound
• Annotations: Expert pixel-wise segmentation masks for every image
• Classification: Parenchymal-inhomogeneity labels for the first 860 images
• License: CC BY-NC-SA 4.0

The dataset card is available on Hugging Face. Download access is provided through the Ditto platform.

Inference with our Pre-trained Model

Pre-trained Cond-UNet Attention weights and a Transformers-compatible image-segmentation pipeline are available on Hugging Face.

from transformers import pipeline

segmenter = pipeline(
    "image-segmentation",
    model="AImageLab-Zip/US_Cond-UNet",
    trust_remote_code=True,
)
result = segmenter("ultrasound.png", organ_id=4)

Pass the corresponding organ_id when the organ is known. If it is omitted, the model uses the unknown-organ token (-1).

Organ	organ_id
Appendix	0
Breast	1
Cardiac	2
Thyroid	3
Fetal / Fetal HC	4
Kidney	5
Liver	6
Testicle	7
Unknown	-1

Dataset Examples

Representative TesticulUS ultrasound images with segmentation-mask overlays:

Citation

If you use the model or dataset for segmentation, please cite:

@inproceedings{morelli2026new,
  title={A New Multicenter Testicular US Dataset and a Lightweight Cond-UNet for Generalization in US Segmentation},
  author={Morelli, Nicola and Marchesini, Kevin and Santi, Daniele and Grana, Costantino and Bolelli, Federico and others},
  booktitle={Proceedings of the British Machine Vision Conference},
  year={2026}
}