CSUN Research Demo

Fair & Explainable
Skin Cancer Detection

A demonstration of the Hybrid DenseNet121–Vision Transformer architecture for dermoscopic image analysis. Trained on ISIC 2020 dataset with AUC of 0.814, achieving equitable performance across all skin tone categories.

0.814
AUC Score
74.0%
Validation Acc.
0.735
Precision
0.75
Recall
Processing Pipeline
🖼️
Input Image
224×224 px RGB
→
✂️
Hair Removal
Morphological filter
→
🌟
CLAHE
Contrast enhance
→
🧠
DenseNet121
Local features
→
⚡
Transformer
Global context
→
📊
Prediction
Sigmoid output

Image Input

🩺
Drop dermoscopic image here
or click to browse · JPG, PNG, WEBP
Sample Cases
🔬
No image loaded
Grad-CAM Visualization

Model Output

Diagnosis
Awaiting Input
—
Malignant —
Benign —
Model
Hybrid DenseNet-ViT
Skin Tone
—
Confidence
—
AUC (Global)
0.814
Skin Tone Fairness Evaluation
Light Skin
0.751
AUC Score
Medium Skin
0.854
AUC Score
Dark Skin
0.941
AUC Score
Baseline Model Comparison
⚠️
Class Collapse Observed. Several baseline models failed to learn a balanced decision boundary due to class imbalance in the dataset. ResNet50 and MobileNetV3 predicted benign for almost every image (Precision = 1.0, Recall = 0.01), while ConvNeXtTiny and Pure ViT predicted malignant for nearly everything (Recall = 1.0, Precision = 0.50). The hybrid model is the only one that achieved genuinely balanced Precision and Recall.
Model Accuracy Precision Recall F1-Score AUC Behaviour
Hybrid DenseNet-ViT ★ Proposed 74.0% 0.735 0.750 0.742 0.814 ✓ Balanced
DenseNet121 67.0% 0.736 0.530 0.617 0.802 Partial bias
ResNet50 50.5% 1.000 ⚠ 0.010 ⚠ 0.020 0.748 Predicts all benign
MobileNetV3 50.5% 1.000 ⚠ 0.010 ⚠ 0.020 0.736 Predicts all benign
ConvNeXtTiny 50.0% 0.500 ⚠ 1.000 ⚠ 0.667 0.706 Predicts all malignant
Pure ViT 50.0% 0.500 ⚠ 1.000 ⚠ 0.667 0.617 Predicts all malignant
Source: Table 1, Gadde (2025) — Hybrid DenseNet–Transformer for Fair and Explainable Skin Cancer Detection, CSUN. F1 computed as 2·(Precision·Recall)/(Precision+Recall). ⚠ flags indicate class-collapsed predictions.
Guna Yaswanth Gadde · California State University Northridge
Hybrid DenseNet–Transformer for Fair and Explainable Skin Cancer Detection
ISIC 2020 Dataset · 1000 Images · 224×224px

⚠️ For educational/research demonstration only. Not a clinical diagnostic tool.