Fourth year capstone project for Electrical Engineering Degree from Lakehead University. Designed and developed AI model for the detection of cancerous epidermal lesions with real-time image classification on ARM hardware. Utilized HAM10000 image database from the International Skin Imaging Collaboration (ISIC).
Setup and configuration of Ubuntu Linux on Nvidia Jetson Nano with Jupyter Lab server environment and Python kernels configured with PyTorch and TorchVision libraries. MATLAB Deep Network Designer utilized for model structure planning and parameter selection. Managed private GitHub for collaboration and version control. Model converted to open standard ONNX for cross-platform compatibility.
Multiple trained AI model architectures tested, with modifications utilizing specialized optimization techniques for further performance improvements. Explored ResNet18 network topology as well as custom topologies for comparative analysis.
Maximum achieved model accuracy of 88% and applied CNN optimization techniques realized a 40% reduction in hardware resource utilization. Final prototype capable of real-time processing and image classification of epidermal lesions with camera input to Jetson Nano. Standalone unit running as headless Linux embedded server secured with GUI internet accessible from any standard web browser.