← All workMedical imaging / cGAN
Resolving the PatchGAN Paradox
A four-model discriminator ablation of conditional GANs for MRI tumour segmentation on BraTS 2020, linking checkerboard artefacts to kernel and stride.
- Stack
- PyTorch · cGAN · U-Net · BraTS 2020
- Key figure
- 0.958 — SSIM
DG Labs
PatchGAN Paradox
0.958SSIM
- Four-model discriminator ablation
- Checkerboard artefacts linked to kernel/stride
- Best: 4×4 stride-2 PatchGAN — PSNR 25.96 dB
PyTorch · cGAN · U-Net · BraTS 2020
01
The problem
Conditional GANs segment brain tumours from MRI well, but their outputs carry checkerboard artefacts — and it is not obvious which part of the discriminator design is responsible.
02
The approach
Ran a four-model discriminator ablation of cGANs with a U-Net generator on BraTS 2020, varying the discriminator's kernel and stride, and traced the artefacts back to those choices.
03
The outcome
A 4×4, stride-2 PatchGAN scored best: PSNR 25.96 dB and SSIM 0.958. The artefacts were linked directly to kernel/stride configuration.
Of note
- ✦ Four-model discriminator ablation
- ✦ Checkerboard artefacts linked to kernel/stride
- ✦ Best: 4×4 stride-2 PatchGAN — PSNR 25.96 dB
- ✦ SSIM 0.958 on BraTS 2020