Quantization-Aware Neuromorphic Architecture for Skin Lesion Classification on Resource-Constrained Devices

Technical Summary: Quantization-Aware Neuromorphic Architecture for Skin Lesion Classification

Problem Statement
On-device skin lesion analysis faces significant barriers due to the high compute and energy costs of conventional Convolutional Neural Network (CNN) inference and the challenges of calibrating models under clinical data shifts. While Spiking Neural Networks (SNNs) and neuromorphic processors (e.g., BrainChip Akida) offer event-driven, sparse computation suitable for resource-constrained environments, practical deployment is hindered by brittle CNN-to-SNN conversion processes. Standard CNN components, such as Batch Normalization, global pooling, and unbounded activations, often fail to convert cleanly into integer spiking graphs, leading to unsupported operators, quantization distortion, and significant accuracy degradation—particularly on small, imbalanced medical datasets where rare lesion cues may be suppressed.

Methodology
The authors propose QANA (Quantization-Aware Neuromorphic Architecture), an end-to-end pipeline designed to ensure conversion-stable neuromorphic execution. The framework consists of three primary stages:

1. Data Preprocessing and Augmentation:

   • Images are deterministically resized to 64×64, padded, and normalized without using saliency models or external detectors.
   • To address class imbalance, the authors employ Synthetic Minority Oversampling (SMOTE) strictly within the learned embedding space. This involves standardizing embeddings, applying PCA whitening, and generating synthetic feature vectors via convex interpolation and local perturbation. These synthetic vectors are injected at the penultimate layer to train the classifier head without altering raw pixel data.

2. Quantization-Aware Network Architecture:

   • Backbone: QANA utilizes a hierarchical design with four stages. Each stage employs Ghost blocks to expand representational diversity under a tight FLOP budget, followed by Batch Normalization (BN) and ReLU6 to bound activations.
   • Attention Mechanisms: The architecture integrates a Spatially-Aware Efficient Channel Attention (SA-ECA) module and a Squeeze-and-Excitation (SE) block. These are designed to be "spike-compatible," meaning their operations (global reduction, channel gating) are explicitly represented as quantized integer operations during conversion.
   • Spike-Compatible Transformation: A transformation module applies bounded activation (clipping values to [0, 1]) and a separable convolution to align activation ranges with spike encoding, preventing saturation under 8-bit affine quantization.
   • Quantized Projection: The final layer maps the compact representation to class logits in a form directly consumable by the CNN-to-SNN converter.

3. CNN-to-SNN Conversion and Deployment:

   • The pipeline enforces an all-integer executable graph on the Akida platform. This involves static reparameterization (folding BN into preceding convolutions), uniform affine 8-bit quantization, and deterministic operator lowering.
   • Operations like global average pooling and fully connected layers are lowered to integer kernels. Residual additions are handled by aligning affine scales to prevent overflow.
   • Deployment: The converted SNN is deployed on BrainChip Akida. Inference is event-driven, using integrate-and-fire neurons. A lightweight readout calibration step adjusts spike thresholds and integration windows using validation data, enabling adaptation without full model retraining or data offloading.

Key Contributions

• QANA Architecture: A conversion-oriented design that replaces fragile components with spike-compatible, bounded-activation blocks and lightweight attention mechanisms, preserving discriminative features under 8-bit constraints.
• Graph-Preserving Conversion Pipeline: A method for Akida that integrates BN reparameterization, integer quantization, and deterministic operator lowering to ensure a fully integer-executable spiking graph.
• Lightweight Calibration Procedure: A deployment strategy that calibrates spike thresholds and readout layers using validation data, allowing for post-conversion adaptation without full retraining.

Experimental Results
The system was evaluated on the HAM10000 benchmark and a proprietary clinical dataset from Nanjing Drum Tower Hospital, deployed on the BrainChip Akida AKD1000.

• Accuracy: On HAM10000, QANA achieved 91.6% Top-1 accuracy and 91.0% macro F1, outperforming the strongest converted SNN baseline (CNN Ensemble) by 3.5 points in accuracy and 12.1 points in macro F1. On the clinical dataset, it achieved 90.8% Top-1 accuracy and 81.7% macro F1, improving the strongest baseline by 3.2 and 3.6 points, respectively.
• Efficiency: On Akida, QANA processes an image in 1.5 ms with 1.7 mJ of energy. Compared to the strongest size-matched Akida baseline (GhostNet), QANA improved accuracy by 4.1 points, macro F1 by 10.0 points, reduced latency by 21.1%, and reduced energy consumption by 22.7%.
• Conversion Stability: The full QANA model demonstrated minimal conversion-induced degradation, with only a 0.4 percentage point drop in Top-1 accuracy and 0.4 points in macro F1 when moving from the quantized CNN to the SNN, significantly outperforming ablated variants.

Significance and Claims
The paper claims that QANA addresses the critical gap between high-performance dermatological AI and the practical constraints of edge deployment. By ensuring conversion stability and supporting lightweight readout calibration, the framework enables real-time, energy-efficient, and privacy-preserving skin lesion classification on neuromorphic hardware. The results suggest that it is possible to maintain high diagnostic accuracy on imbalanced medical datasets while adhering to strict latency and energy budgets, making it a viable solution for portable decision-support systems in home and remote clinical settings. The authors emphasize that their approach avoids the security risks and regulatory hurdles associated with cloud-centric data transfer.