Diffusers
TensorBoard
controlnet
stable-diffusion
medical-imaging
retinal-imaging
diabetic-retinopathy
counterfactual
Instructions to use KylianSu/vessel-bezier-retinal-weights with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use KylianSu/vessel-bezier-retinal-weights with Diffusers:
pip install -U diffusers transformers accelerate
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline controlnet = ControlNetModel.from_pretrained("KylianSu/vessel-bezier-retinal-weights") pipe = StableDiffusionControlNetPipeline.from_pretrained( "fill-in-base-model", controlnet=controlnet ) - Notebooks
- Google Colab
- Kaggle
§5.3 Prompt-fixed 对比表(EfficientNet-B2 @ 384, leak-free test n=13,237)
格式:Test AUC / Sensitivity / F1
全量训练(n_train_real=114,209)
| Config | Old (prompt-bug) AUC / Sens / F1 | New (prompt-fixed) AUC / Sens / F1 | Δ (new − old) AUC / Sens / F1 |
|---|---|---|---|
| Baseline (real only) | 0.9612 / 0.8858 / 0.9047 | 0.9612 / 0.8858 / 0.9047 | (baseline 复用) |
| G2 + 400 synth | 0.9594 / 0.8749 / 0.9058 | 0.9609 / 0.8492 / 0.9085 | +0.0015 / -0.0257 / +0.0027 |
| G2 + 4k synth | 0.9625 / 0.8868 / 0.9105 | 0.9553 / 0.8505 / 0.9043 | -0.0072 / -0.0363 / -0.0062 |
| G3 + 400 synth | 0.9592 / 0.8920 / 0.9081 | 0.9614 / 0.8358 / 0.9020 | +0.0022 / -0.0562 / -0.0061 |
| G3 + 4k synth | 0.9615 / 0.8513 / 0.9074 | 0.9589 / 0.8806 / 0.9049 | -0.0026 / +0.0293 / -0.0025 |
小样本 10%(n_train_real=11,420)
| Config | Old (prompt-bug) AUC / Sens / F1 | New (prompt-fixed) AUC / Sens / F1 | Δ (new − old) AUC / Sens / F1 |
|---|---|---|---|
| Baseline 10% real | 0.9393 / 0.8295 / 0.8693 | 0.9393 / 0.8295 / 0.8693 | (baseline 复用) |
| G2 + 400 synth (10%) | 0.9377 / 0.8303 / 0.8681 | 0.9351 / 0.8257 / 0.8694 | -0.0026 / -0.0046 / +0.0013 |
| G2 + 4k synth (10%) | 0.9341 / 0.8238 / 0.8684 | 0.9376 / 0.8141 / 0.8707 | +0.0035 / -0.0097 / +0.0023 |
| G3 + 400 synth (10%) | 0.9375 / 0.8261 / 0.8699 | 0.9322 / 0.8139 / 0.8629 | -0.0053 / -0.0122 / -0.0070 |
| G3 + 4k synth (10%) | 0.9366 / 0.8142 / 0.8696 | 0.9349 / 0.8108 / 0.8656 | -0.0017 / -0.0034 / -0.0040 |
关键变化(仅看 Test AUC 提升)
| 配置 | 旧 ΔAUC vs baseline | 新 ΔAUC vs baseline | 新-旧 ΔAUC |
|---|---|---|---|
| G2 + 400 synth | -0.0018 | -0.0003 | +0.0015 |
| G2 + 4k synth | +0.0013 | -0.0059 | -0.0072 |
| G3 + 400 synth | -0.0020 | +0.0002 | +0.0022 |
| G3 + 4k synth | +0.0003 | -0.0023 | -0.0026 |
| G2 + 400 synth (10%) | -0.0016 | -0.0042 | -0.0026 |
| G2 + 4k synth (10%) | -0.0052 | -0.0017 | +0.0035 |
| G3 + 400 synth (10%) | -0.0018 | -0.0071 | -0.0053 |
| G3 + 4k synth (10%) | -0.0027 | -0.0044 | -0.0017 |