[UIUC CS598 LHO/DL4H, fanz6] ISIC2018 + PH2 + Diffusion + Training + Eval

[ffanzhang]

Name: Fan Zhang
NetID: fanz6
Email: fanz6@illinois.edu
Type of contribution: Dataset + Task
Paper: https://proceedings.mlr.press/v287/jin25b.html

Implementation:

Datasets

• ISIC2018 Dataset (https://challenge.isic-archive.com/data/#2018)
• An annotation Dataset based on ISIC2018 (https://github.com/alceubissoto/debiasing-skin)
• PH2 (https://www.fc.up.pt/addi/ph2%20database.html), download link broken, used (https://github.com/vikaschouhan/PH2-dataset)

Files of Interest

• Reference Code (https://github.com/QixuanJin99/dermoscopic_artifacts)

examples/isic2018_train_resnet50.py

• Trains a ResNet-50 melanoma classifier on ISIC 2018 with 5-fold cross-validation
• Supports multiple image modes (whole, lesion, background, bbox) and frequency filtering (high/low)
• Saves per-fold checkpoints used by downstream evaluation scripts

examples/isic2018_artifacts_train_lora_sd15.py

• Trains a DreamBooth+LoRA model (SD 1.5) per artifact type on ISIC 2018 artifact-labeled images
• Matches paper §3.3: rank-64 LoRA on UNet + CLIP text encoder, rare token per artifact, prior preservation
• Artifacts: dark_corner, hair, gel_bubble, ruler, ink, patches
• Saves LoRA adapters to ~/lora_checkpoints/{artifact}/

examples/ph2_train_resnet50.py

• Trains a ResNet-50 melanoma classifier on PH2 with 5-fold cross-validation
• Used as the PH2-trained baseline in Table 4 evaluation

examples/ph2_diffusion_sd.py

• Generates artifact-augmented PH2 images using SD inpainting + trained LoRA adapters
• Programmatic seeds (hair strands, ink marks, ruler strips, etc.) guide SD inpainting per artifact
• Outputs ~/ph2_augmented/{artifact}/{image_id}.jpg and augmented_metadata.csv

examples/ph2_artifacts_test_resnet50.py

• Replicates Table 4 from Jin et al. (CHIL 2025): evaluates ResNet-50 classifiers (trained on PH2 or ISIC) on diffusion-augmented PH2 images
• Reports AUROC per artifact condition across 5 folds for both classifier sources

[jhnwu3]

None of the examples here use your pyhealth loaders.

[jhnwu3]

Nothing PyHealth related here.

[jhnwu3]

Tldr; The PyHealth loaders/tasks look pretty good. But, it seems none of the examples here use it to do anything/training. For reference SampleDataset is effectively the same thing as a PyTorchDataset class but with precached tensors.

[ffanzhang]

None of the examples here use your pyhealth loaders.

I will refactor to use pyhealth loaders