Neuromorphic AI chip efficiently diagnoses skin lesions on pocket-sized devices

Deploying neural networks for medical imaging on battery-powered edge devices remains a stubborn technical challenge. Researchers have now introduced QANA, a neuromorphic architecture that bridges the gap between traditional deep learning models and energy-efficient processors specifically designed for event-driven computation.

The efficiency problem#

Conventional neural networks burn through computational resources when deployed on mobile or embedded hardware. Neuromorphic processors—which mimic biological spike-based processing—promise to cut energy consumption and latency dramatically, but converting standard CNNs to spiking neural networks (SNNs) often fails due to incompatible operations, quantization distortion, and accuracy loss, especially when training data is imbalanced.

How QANA works#

The proposed system redesigns the CNN backbone to play nicely with neuromorphic conversion. Key improvements include:

• Controlled activations: Bounding intermediate values prevents distortion during the CNN-to-SNN transition
• Aligned quantization: Normalization layers are structured to work with low-bit precision without sacrificing accuracy
• SNN-ready components: Ghost-based feature generation and attention modules are swapped in to replace operations that don't survive conversion
• Lightweight calibration: The quantized output head supports retuning on new clinical data without full retraining

Benchmark results#

On the HAM10000 skin lesion dataset, QANA achieves 91.6% top-1 accuracy and 91.0% macro F1 score, outpacing the next-best converted SNN by 3.5 percentage points and 12.1 points in F1. On a separate clinical dataset, the model hits 90.8% accuracy and 81.7% macro F1, beating the closest baseline by 3.2 and 3.6 points respectively.

On BrainChip's Akida neuromorphic processor, the system runs individual images in just 1.5 milliseconds while burning only 1.7 millijoules per inference—21% faster and 23% more efficient than comparable baselines, with 4.1 and 10.0 point improvements in accuracy and F1.

Why it matters#

The work addresses a real clinical need: real-time skin cancer screening on handheld devices without draining batteries or requiring internet connectivity. By solving the conversion bottleneck, QANA opens a path toward on-device medical diagnostics that scales to resource-constrained settings.