You can deploy this model and start earning money today!

yushangjinghong/
t1-mri-denoising

Name: t1-mri-denoising
Availability: InStock

Views

No views yet

T1 MRI Denoising — 2.5D Residual U-Net with SE Attention

Author: 雷钧翔 | Student ID: 23307140057

2.5D Residual U-Net for T1 MRI denoising (noisy T1 → clean T1), with Squeeze-and-Excitation channel attention and composite loss (L1 + SSIM + Gradient).

Models

File	Params	PSNR	SSIM
`small/best.pt`	31M	44.99 dB	0.9907
`large/best.pt`	193.7M	45.14 dB	0.9908

Noisy baseline: PSNR = 43.34 dB, SSIM = 0.9664

Architecture

Input: 5 adjacent axial slices stacked as channels (2.5D), 192×192
Backbone: Residual U-Net with InstanceNorm + SiLU activations
Attention: SE (Squeeze-and-Excitation) after each ConvBlock
Output: Residual prediction → pred_clean = clamp(noisy_center + residual, 0, 1)

Training

Loss: L1 + 0.1×(1−SSIM) + 0.05×GradLoss
SSIM: 11×11 Gaussian kernel (σ=1.5)
Optimizer: AdamW (lr=2e-4, wd=1e-4)
Scheduler: CosineAnnealingLR (η_min=1e-6)
AMP fp16 + gradient clipping (max_norm=1.0)

Usage

python

import torch
checkpoint = torch.load("large/best.pt", map_location="cpu")
# checkpoint contains: model_state_dict, args, epoch, best_val_psnr

See GitHub repo for full training code.

yushangjinghong/t1-mri-denoising