EmmaEgbo/EdgeDerm-Quantized

Offline-ready skin lesion classifier optimized for low-resource edge devices. Uses static quantization to reduce ResNet-18 size by 75% with <0.1% accuracy loss. # Edge-Optimized Dermatological Diagnosis (Quantized ResNet-18) ## Project Overview This project addresses the challenge of deploying medical diagnostic tools in low-connectivity environments (e.g., rural clinics in West Africa). By implementing **Post-Training Static Quantization (PTQ)** on a ResNet-18 architecture, I developed a skin lesion classifier that retains diagnostic accuracy while significantly reducing computational requirements. ## Key Engineering Outcomes | Metric | Original (FP32) | Quantized (Int8) | Improvement | | :--- | :---: | :---: | :---: | | **Model Size** | 44.77 MB | 11.31 MB | **74.8% Reduction** | | **Inference Latency** | 21.79 ms | 15.69 ms | **1.4x Speedup** | | **Accuracy** | 72.87% | 72.82% | **-0.05% (Lossless)** | *Benchmarks run on CPU to simulate edge device constraints.* ## Technical Stack * **Framework:** PyTorch (Torch.nn, Torch.quantization) * **Architecture:** Custom ResNet-18 (modified for 28x28 input) * **Dataset:** MedMNIST v2 (DermaMNIST) - 10,015 images, 7 classes. * **Optimization:** Layer Fusion (Conv+BN+ReLU) & Static Quantization (fbgemm). ## Methodology 1. **Architecture Design:** Adapted a ResNet-18 backbone by replacing the initial 7x7 convolution with a 3x3 kernel to handle low-resolution biomedical imagery. 2. **Training:** Trained on DermaMNIST using CrossEntropyLoss and Adam optimizer. 3. **Optimization Pipeline:** * **Fusion:** Merged Convolution, BatchNorm, and ReLU layers to reduce memory access overhead. * **Calibration:** Used a representative dataset to determine dynamic ranges for activation quantization. * **Conversion:** Mapped weights from Float32 to Int8. ## Future Scope * Deployment to Android via **ONNX Runtime**. * Integration with offline-first mobile application.