This repository contains the official implementation of UltraSharp, a geometry-aware transformer architecture for single-image super-resolution of clinical ultrasound data. The method is grounded in Beltrami operator theory and Riemannian geometry, treating the image domain as a curved manifold whose local metric is driven by tissue structure and physical acquisition properties.
Accepted as Oral Presentation: IEEE ISBI 2026
Prabhav Sanga1, Jaskaran Singh2, Tapabrata Chakraborti1,3
1University College London 2University of Nottingham 3The Alan Turing Institute, UK
Figure 1. Qualitative comparison on CAMUS (echocardiography) and BUSI (breast ultrasound) at 4x super-resolution. UltraSharp recovers fine anatomical detail and preserves clinically relevant speckle texture.
Figure 2. Extended comparisons across scale factors (x2, x4, x8) and ablation study isolating the contribution of each architectural component.
Standard image super-resolution methods disregard the physical formation process of ultrasound and the Riemannian structure of tissue boundaries. UltraSharp addresses both limitations through three contributions:
Beltrami Positional Encoding (BPE). Rather than using fixed sinusoidal or learned position tokens, BPE diffuses anchor impulses under Beltrami flow on the image manifold. The resulting positional channels encode geodesic distance from landmarks in a way that is anisotropic and structure-aware, guided by the inverse Riemannian metric tensor computed from the local image gradient distribution.
Anisotropic Geodesic Attention (AGA). Self-attention within local windows is modulated by a geodesic proximity term derived from the Beltrami embeddings. Concretely, the log of a Riemannian proximity weight is added to the attention logits before softmax, suppressing cross-boundary mixing and concentrating information exchange along anatomical structures.
Physics-Constrained Fusion (PCF). The decoder couples the feature branch with a physics-simulation branch that applies learnable anisotropic Gaussian PSFs parameterised by the estimated Riemannian metric. This enforces cycle consistency between the generated high-resolution image and the physical degradation model, acting as a structural regulariser.
The composite loss function combines pixel-level L1 reconstruction, structural similarity, Beltrami geometric regularisation, Rayleigh speckle KL divergence, and physics cycle-consistency.
Four capacity configurations are provided. All share the same architecture and differ only in width and depth.
| Variant | Embedding dim | Heads | Blocks | Parameters |
|---|---|---|---|---|
| ultrasharp-t | 32 | 4 | [1, 1, 1] | ~5M |
| ultrasharp-s | 48 | 6 | [2, 2, 1] | ~11M |
| ultrasharp-b | 64 | 8 | [2, 2, 2] | ~22M (paper default) |
| ultrasharp-l | 96 | 12 | [3, 3, 3] | ~45M |
Pre-trained weights will be released after ISBI 2026. Refer to checkpoints/CHECKPOINTS.md for the expected file layout.
from models.builder import build_ultrasharp
model = build_ultrasharp("ultrasharp-b", scale=4)git clone https://github.com/YourOrg/UltraSharp.git
cd UltraSharp
pip install -r requirements.txtPython 3.8 or later and PyTorch 2.0 or later are required. A CUDA-capable GPU is strongly recommended.
Experiments were conducted on four publicly available ultrasound datasets.
| Dataset | Modality | Training / Val / Test | Source |
|---|---|---|---|
| CAMUS | Echocardiography | 450 / 50 / 50 | CREATIS challenge |
| EchoNet-Dynamic | Echocardiography (video) | 7,465 / 1,288 / 1,277 | echonet.github.io |
| BUSI | Breast ultrasound | 547 / 100 / 80 | Kaggle |
| HC18 | Fetal head circumference | 999 / 100 / 100 | Grand Challenge |
Place images in a flat directory structure. Low-resolution inputs are generated on-the-fly during training by the physics-aware degradation pipeline (data/synthesis.py); no pre-computed LR images are required.
data/
train/ *.png (or .jpg)
val/ *.png
test/ *.png
Dataset-specific loaders with ground-truth segmentation masks (required for CNR and sSNR evaluation) will be released together with the pre-trained checkpoints.
Full training code and pre-trained checkpoints will be released after ISBI 2026. The model architecture, loss functions, and degradation pipeline are fully provided.
python scripts/train.py \
--model ultrasharp-b \
--data_dir /path/to/data/train \
--scale 4 \
--epochs 200 \
--batch_size 8 \
--augmentResults at 4x super-resolution on the CAMUS test set. Ultrasound-specific metrics Contrast-to-Noise Ratio (CNR) and speckle Signal-to-Noise Ratio (sSNR) are computed in addition to standard SR metrics.
| Method | PSNR (dB) | SSIM | LPIPS | CNR | sSNR |
|---|---|---|---|---|---|
| Bicubic | 27.3 | 0.741 | 0.312 | 1.2 | 3.1 |
| SRCNN | 29.1 | 0.802 | 0.261 | 1.6 | 3.8 |
| EDSR | 31.8 | 0.851 | 0.198 | 2.1 | 4.7 |
| SwinIR | 33.2 | 0.878 | 0.162 | 2.4 | 5.3 |
| UltraSharp (Ours) | 35.1 | 0.913 | 0.118 | 3.1 | 6.8 |
UltraSharp/
models/
ultrasharp.py Main U-Net architecture
transformer_block.py Beltrami Transformer Block (BTB)
attention.py Anisotropic Geodesic Attention (AGA)
bpe.py Beltrami Positional Encoding
pcm.py Physics-Constrained Fusion decoder
builder.py Model factory (four variants: T / S / B / L)
data/
synthesis.py Physics-aware degradation pipeline
dataset.py Dataset base class [released post-conference]
transforms.py Augmentation utilities
losses/
losses.py Beltrami, Speckle, Physics-cycle, L1 losses
utils/
structure_tensor.py Riemannian structure tensor and Beltrami metric
metrics.py PSNR, SSIM, LPIPS, CNR, sSNR
scripts/
train.py Training entry point [full release post-ISBI 2026]
assets/ Figures from the paper
requirements.txt
README.md
@inproceedings{sanga2026ultrasharp,
author = {Sanga, Prabhav and Singh, Jaskaran and Chakraborti, Tapabrata},
title = {UltraSharp: Beltrami Transformers for Ultrasound Super-Resolution},
booktitle = {Proceedings of the IEEE International Symposium on Biomedical Imaging (ISBI)},
year = {2026},
address = {London, UK},
note = {Oral Presentation}
}This project is released under the MIT License. See LICENSE for details.

