QANA: Neuromorphic Skin Lesion Classification for Edge Devices

A recent preprint on arXiv (2507.15958v5) presents QANA, an architecture that is quantization-aware and tailored for classifying skin lesions on devices with limited resources. This study tackles the difficulties faced in on-device medical imaging, where traditional CNN inference struggles due to computational and energy limitations, necessitating lightweight calibration amid shifts in clinical data. While neuromorphic processors enable event-driven sparse computations, their effective use is often obstructed by issues in converting CNNs to SNNs, such as unsupported operators, quantization errors, and accuracy loss due to class imbalance. QANA introduces a quantization-aware CNN backbone within a comprehensive pipeline to ensure stable neuromorphic execution. It enhances conversion reliability by constraining intermediate activations, synchronizing normalization with low-bit quantization, and substituting fragile components with spike-compatible alternatives. Efficient representation is facilitated through Ghost-based feature generation, spatially-aware channel attention, and quantizable squeeze-and-excitation modules. This research is significant at the crossroads of AI, healthcare, and edge computing, paving the way for effective skin lesion analysis on low-power devices.