GStocks123
Regular
Worth the read!
The deployment platform, BrainChip’s Akida (ASX:BRN;OTC:BRCHF) neuromorphic processor, natively supports SNN operations and enables on-chip incremental learning. This capability allows the system to adapt to new patient data without requiring complete model retraining, aligning with clinical workflows that continuously encounter new diagnostic cases.
The Squeeze-and-Excitation blocks implement adaptive channel weighting through a two-stage bottleneck mechanism, providing additional regularization particularly beneficial for small, imbalanced medical datasets. The quantized output projection produces SNN-ready outputs that can be directly processed by neuromorphic hardware without additional conversion steps.
On HAM10000, QANA achieved 91.6% Top-1 accuracy and 82.4% macro F1 score, maintaining comparable performance on the clinical dataset with 90.8% accuracy and 81.7% macro F1 score. These results demonstrate consistent performance across both standardized benchmarks and clinical practice conditions.
The system showed balanced performance across all seven diagnostic categories, including minority classes such as dermatofibroma and vascular lesions. Notably, melanoma detection achieved 95.6% precision and 93.3% recall, critical metrics for this potentially life-threatening condition.
Comparative analysis against state-of-the-art architectures converted to SNNs showed QANA’s superior performance across all metrics. While conventional architectures experienced accuracy drops of 3–7% after SNN conversion, QANA maintained high accuracy through its quantization-aware design principles.
Neuromorphic Computing Advantages
Moving to Spiking Neural Networks (SNNs) signifies a core alteration from standard neural processing. In contrast to standard neural networks that process continuous values, SNNs utilize discrete spike events for information encoding and transmission. This method yields sparse, event-driven computation that greatly lessens energy use while preserving diagnostic accuracy.The deployment platform, BrainChip’s Akida (ASX:BRN;OTC:BRCHF) neuromorphic processor, natively supports SNN operations and enables on-chip incremental learning. This capability allows the system to adapt to new patient data without requiring complete model retraining, aligning with clinical workflows that continuously encounter new diagnostic cases.
Spike-Compatible Architecture Design
The architecture incorporates specific design elements that facilitate seamless conversion from standard neural networks to spike-based processing. The spike-compatible feature transformation module employs separable convolutions with quantization-aware normalization, ensuring all activations remain within bounds suitable for spike encoding while preserving diagnostic information.The Squeeze-and-Excitation blocks implement adaptive channel weighting through a two-stage bottleneck mechanism, providing additional regularization particularly beneficial for small, imbalanced medical datasets. The quantized output projection produces SNN-ready outputs that can be directly processed by neuromorphic hardware without additional conversion steps.
Comprehensive Performance Validation
Experimental validation was conducted on both the HAM10000 public benchmark dataset and a real-world clinical dataset from Hospital Sultanah Bahiyah, Malaysia. The clinical dataset comprised 3,162 dermatoscopic images from 1,235 patients, providing real-world validation beyond standard benchmarks.On HAM10000, QANA achieved 91.6% Top-1 accuracy and 82.4% macro F1 score, maintaining comparable performance on the clinical dataset with 90.8% accuracy and 81.7% macro F1 score. These results demonstrate consistent performance across both standardized benchmarks and clinical practice conditions.
The system showed balanced performance across all seven diagnostic categories, including minority classes such as dermatofibroma and vascular lesions. Notably, melanoma detection achieved 95.6% precision and 93.3% recall, critical metrics for this potentially life-threatening condition.
Hardware Performance and Energy Analysis
When deployed on Akida hardware, the system delivers 1.5 millisecond inference latency and consumes only 1.7 millijoules of energy per image. These figures represent reductions of over 94.6% in inference latency and 98.6% in energy consumption compared to GPU-based CNN implementations.Comparative analysis against state-of-the-art architectures converted to SNNs showed QANA’s superior performance across all metrics. While conventional architectures experienced accuracy drops of 3–7% after SNN conversion, QANA maintained high accuracy through its quantization-aware design principles.
. 
links to an excellent blog post written by our Chief Development Officer Jonathan Tapson titled 
