Liam Liu
Overview

Medical imaging · Generative models

SCAM-GAN: Attention-Augmented CT-to-MR Synthesis

A generative model that synthesizes brain MR images from CT scans without paired training data.

Published · JPCS 2646
System architecture01 — 05
01CT / MR Datasets
02Preprocessing
03Attention CycleGAN
04MAE / PSNR / SSIM
05Ablation Study

My scope

The boundary I owned.

I designed the attention-augmented CycleGAN, built the PyTorch training and evaluation pipeline, and ran the ablation studies behind the published paper.

  • Python
  • PyTorch
  • CycleGAN
  • CBAM
  • NumPy
  • Computer Vision

Engineering evidence

  1. Engineered reusable PyTorch pipelines for preprocessing, augmentation, model training, checkpointing, evaluation, and reproducible experiments.
  2. Designed an attention-augmented CycleGAN with spatial and channel attention (CBAM) for unpaired CT-to-MR image translation.
  3. Evaluated paired and unpaired baselines with MAE, PSNR, and SSIM alongside ablation studies, qualitative inspection, and anatomical error analysis.
OutcomePeer-reviewed results with sharper anatomical fidelity than unattended baselines.