PatchGuard: Deep Learning for Early Detection of Skin Cancer using Neural ODE-Enhanced ResNet18

Data Preparation• Prepare ISICC 2020 Dataset• Resize and normalize images• Simulate corruption• Apply RandomErasing = 0.5 (scale 0.02–0.2) on hbsc• GaussianBlur (kernel size 5)

Baseline Model Training• Train ResNet- 18 on clean ISIC images• Target validation accuracy of 86% or higher

Hybrid Model Training• Calculate loss as ResNet, Neural ODE, and Masking• Target robust accuracy 70% under FGSM

PatchGuard++ Core ArchitectureReplace ResNet classifier with Neural ODE• class ODEBlock(...)

Adversarial Testing on Baseline• Using Foolbox FGSM Attacks with Misclassification criterio to generate adversarial images• Measure accuracy drop 70 > 45%

Visualization & Explainability• Use Grad-CAM to generate attention maps• Demonstrating model focussing on multiple lesion regions

Robustness Benchmarking• Evaluation using ISIC-C dataset containing 10 corruption types (like Mon Biur. Shadows, opex,)• Comparing drop in accuracy with the baseline

Ablation Studies• Ablation 1: Remove Noui. ODE and replaced with CNN (expecting drop 10-15% under noise)• Ablation 2: Remove dynamic masking, expect overfitting to unmasked patches• Visualization & Explainability• Use Grad-CAM to generate attention maps on multiple lesion regions

advi