SkinGenBench: Generative Model and Preprocessing Effects for Synthetic Dermoscopic Augmentation in Melanoma Diagnosis

📢 Official PyTorch implementation of SkinGenBench: Generative Model and Preprocessing Effects for Synthetic Dermoscopic Augmentation in Melanoma Diagnosis
N. A. Adarsh Pritam, Jeba Shiney O, Sanyam Jain
[Paper] | [Code]

---

🌟 Highlights

• First Systematic Benchmark: Controlled evaluation of preprocessing complexity (basic geometric vs. advanced artifact removal) across GANs (StyleGAN2-ADA) and Diffusion Models (DDPM) for dermoscopic melanoma synthesis.
• Architecture > Preprocessing: Demonstrates that generative model choice has stronger impact than preprocessing complexity on both image fidelity and diagnostic utility.
• StyleGAN2-ADA Superiority: Achieves lowest FID (≈65.5) and KID (≈0.05) with better class anchoring, while diffusion models produce higher variance at the cost of perceptual fidelity.
• Significant Clinical Impact: Synthetic augmentation delivers 8-15% absolute melanoma F1-score improvements, with ViT-B/16 reaching F1≈0.88 and ROC-AUC≈0.98 (≈14% improvement over baselines).
• Reproducible Framework: Unified assessment combining generative metrics (FID, IS, KID), downstream performance across five architectures (CNNs and transformers), and interpretability analysis via Grad-CAM on 14,116 dermoscopic images.

📊 Dataset

Overall experimental design showing dual preprocessing pipelines, generative model training, synthetic data augmentation, and downstream classifier evaluation for melanoma diagnosis.

Table: Overview of curated dermatology dataset used in our study. The dataset combines ISIC 2025 (MLK10k) and HAM10000 sources.

Class	Abbr.	Images	Percentage
Nevus	NV	7,424	52.60%
Basal Cell Carcinoma	BCC	3,026	21.43%
Benign Keratosis-like	BKL	1,637	11.60%
Melanoma	MEL	1,563	11.03%
Squamous Cell Carcinoma	SCC	466	3.34%
Total		14,116	100%

Dataset Sources:

• ISIC 2025 (MLK10k)
• HAM10000

---

🧠 Overview

Figure: General framework of SkinGenBench showing the two preprocessing pipelines (Basic and Advanced), generative model training (StyleGAN2-ADA and DDPM), and evaluation through image quality metrics and downstream classification tasks.

Image Subset Nomenclature:

Source	Basic Preprocessing (BS)	Advanced Preprocessing (AD)
Ground Truth	BS_GT	AD_GT
StyleGAN2-ADA	BS_GN	AD_GN
DDPM	BS_DF	AD_DF
Ground-Truth Aug.	BS_GTA	AD_GTA

---

📄 Publication

SkinGenBench: Generative Model and Preprocessing Effects for Synthetic Dermoscopic Augmentation in Melanoma Diagnosis
N. A. Adarsh Pritam, Jeba Shiney O, Sanyam Jain
Alliance University, Bangalore & Østfold University College, Norway
[GitHub] | [PDF]

---

📑 Table of Contents

• Installation
• Model Zoo
• Training
• Results
• Citation

---

Installation

1. Clone the repository:

   git clone https://github.com/adarsh-crafts/SkinGenBench.git
   cd SkinGenBench

2. Create a virtual environment:

   python3 -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

3. Install PyTorch: Install PyTorch following instructions from PyTorch official site.

requirements.txt

torch>=2.0.0
torchvision>=0.15.0
numpy
opencv-python
matplotlib
scikit-learn
scipy
tqdm
h5py
pandas
Pillow

---

Model Zoo

Pretrained Models for StyleGAN2-ADA and DDPM which were finetuned are available here in table below:

StyleGAN2-ADA Models

Model	Configuration	File
StyleGAN2-ADA	FFHQ 256×256 pretrained	NVIDIA CDN
DDPM	CELEBA-HQ 256×256 pretrained	Hugging Face

---

Training

Train StyleGAN2-ADA, DDPM and the classifiers with the provided configurations in each nested directory.

Training Details:

Configuration	Minimum	Maximum
GPU	NVIDIA RTX 4060 8GB × 1	NVIDIA L4 22GB × 1
RAM	8 GB	22 GB
Input Resolution	256×256×3	256×256×3

---

Results

Figure: t-SNE embeddings showing ground truth (GT), StyleGAN2-ADA (GN), and DDPM (DF) distributions for basic (left) and advanced (right) preprocessing pipelines.

Performance Metrics

Generative Model Quality Metrics (Epoch 1000)

Fréchet Inception Distance (FID) - Lower is better

Model	Basic Pipeline (BS)	Advanced Pipeline (AD)
StyleGAN2-ADA (BSGN/ADGN)	79.36	65.47
DDPM (BSDF/ADDF)	83.04	90.22

Kernel Inception Distance (KID) - Lower is better

Model	Basic Pipeline (BS)	Advanced Pipeline (AD)
StyleGAN2-ADA (BSGN/ADGN)	0.0664	0.0546
DDPM (BSDF/ADDF)	0.0684	0.0772

Inception Score (IS) - Higher is better

Model	Basic Pipeline (BS)	Advanced Pipeline (AD)
StyleGAN2-ADA (BSGN/ADGN)	3.22	2.77
DDPM (BSDF/ADDF)	2.50	2.45

---

Global Classification Performance

Best Results: Pipeline A2 (Basic Preprocessing + StyleGAN2-ADA Augmentation)

Model	Macro-F1	Balanced Acc	MCC	ROC-AUC	Accuracy	Brier Score ↓
ViT-B/16	0.8393	0.8348	0.8515	0.9822	0.8985	0.0302
ResNet-50	0.8393	0.8433	0.8525	0.9802	0.8989	0.0314
VGG-16	0.8181	0.8167	0.8243	0.9774	0.8797	0.0365
EfficientNet-B0	0.7977	0.7871	0.8033	0.9698	0.8657	0.0402
ResNet-18	0.7525	0.7424	0.7594	0.9611	0.8360	0.0479

---

Melanoma (MEL) Classification Performance

Best Results: Pipeline A2 (Basic Preprocessing + StyleGAN2-ADA Augmentation)

Model	MEL F1	Sensitivity	Specificity	Precision	ROC-AUC	PR-AUC	DOR
ViT-B/16	0.8831	0.8564	0.9798	0.9115	0.9802	0.9511	288.94
ResNet-50	0.8663	0.8401	0.9758	0.8941	0.9787	0.9445	211.93
VGG-16	0.8438	0.8108	0.9730	0.8796	0.9729	0.9228	154.56
EfficientNet-B0	0.8126	0.7781	0.9667	0.8503	0.9633	0.9043	101.76
ResNet-18	0.7724	0.7390	0.9576	0.8089	0.9542	0.8774	63.88

Key Melanoma Detection Improvements (A2 vs A1)

• MEL F1-score gains: +8–15% across all architectures
• ViT-B/16: MEL F1 improved from 0.7401 → 0.8831 (+14.3 percentage points)
• ResNet-50: MEL F1 improved from 0.7362 → 0.8663 (+13.0 percentage points)
• All models achieved ROC-AUC > 0.96 for melanoma detection

---

Pipeline Comparison Summary

Pipeline	Description	Best Use Case
A2	Basic preprocessing + StyleGAN2-ADA	Recommended: Best overall performance
A3	Basic preprocessing + DDPM	Good diversity, lower fidelity
B2	Advanced preprocessing + StyleGAN2-ADA	Marginal gains over A2
B3	Advanced preprocessing + DDPM	Lowest performance
A4/B4	Standard augmentation only (no synthetic)	Baseline comparison

Key Finding: Generative architecture choice (GAN vs Diffusion) has a stronger influence on diagnostic performance than preprocessing complexity (Basic vs Advanced).

Figure: Grad-CAM visualizations comparing ResNet-50 and ViT-B/16 across different preprocessing pipelines and generative models. ResNet-50 produces compact, lesion-aligned saliency maps, while ViT-B/16 shows broader attention patterns. Synthetic samples exhibit more irregular activations, with ADDF showing the smoothest, most anatomically coherent results.

---

Citation

If you find this work useful, please cite our paper:

@misc{pritam2025skingenbenchgenerativemodelpreprocessing,
      title={SkinGenBench: Generative Model and Preprocessing Effects for Synthetic Dermoscopic Augmentation in Melanoma Diagnosis}, 
      author={N. A. Adarsh Pritam and Jeba Shiney O and Sanyam Jain},
      year={2025},
      eprint={2512.17585},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2512.17585}, 
}

---