sfczaa/ddpm-derm-c1-seed2

HAM10000 ResNet-18 — C1@585 seed 2

This repository contains the deployment checkpoint best.pt for an educational, non-commercial portfolio demonstration. It is not a medical device and must not be used for diagnosis, treatment, triage, or other clinical decisions.

Model selection

• Architecture: ResNet-18
• Variant: C1@585
• Seed: 2
• Selection rule: highest validation dermatofibroma (df) F1 among the completed C1@585 seeds 0, 1, and 2
• Seed-2 validation df F1: 0.740741
• Test metrics were not used to choose this deployment checkpoint

The matched-585 formal experiment averaged across three seeds produced df F1 0.660 ± 0.042, macro-F1 0.651 ± 0.021, and df recall 0.604 ± 0.029. These fixed-split results are suggestive only and are not a claim of statistical significance, clinical validation, or generalization to other populations.

Input and classes

Images are converted to RGB, resized to 128 × 128, converted to a tensor, and normalized with ImageNet mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225].

Class order:

akiec, bcc, bkl, df, mel, nv, vasc

File integrity

• File: best.pt
• SHA-256: c55134294687a753cff44ac5bac5a4896f85adde1be943697feab08129350752

The demo validates this digest and the checkpoint metadata before serving predictions.

Training data, license, and attribution

The model was trained on the HAM10000 dataset using a fixed lesion-ID group split. HAM10000 Dataset © ViDIR Group, Department of Dermatology, Medical University of Vienna. The ISIC 2018 data page distributes the dataset under CC BY-NC 4.0.

Reference: Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data 5, 180161 (2018). doi:10.1038/sdata.2018.161

Modifications include a fixed train/validation/test split, resizing, normalization, class rebalancing for C1@585, and model training. No endorsement by the original dataset creators or ISIC is implied.

Limitations

The model may be wrong. HAM10000 and the fixed test split do not represent all devices, acquisition settings, skin tones, lesion types, or patient populations. The test set contains few df cases, so per-class estimates are uncertain.