NeuroMorphFusion: A Neuro-Inspired Hybrid Learning Framework for Interpretable Deep Lesion Detection in IoT-Enabled Healthcare Systems

TL;DR

NeuroMorphFusion is a new framework for detecting lesions in medical imaging that combines brain-inspired learning with efficient computation for use in healthcare IoT systems.

Contribution

The novel framework integrates spiking neural networks, morphological attention, and genetic algorithm optimization for interpretable and efficient lesion detection.

Findings

NeuroMorphFusion achieves 98.18% classification accuracy on the IQ-OTHNCCD lung cancer CT dataset.

The framework outperforms VGG16, SqueezeNet, MobileNetV3, and ResNet18 in transparency and efficiency.

Optimization converges within eight minutes on a Jetson Nano, balancing accuracy, latency, and explainability.

Abstract

Integrating deep learning within the Internet of Medical Things (IoMT) has revolutionized automated lesion detection in medical imaging. Yet, maintaining high diagnostic accuracy, interpretability and computational efficiency on resource-limited edge devices remains challenging. To address these gaps, we propose NeuroMorphFusion, a neuro-inspired hybrid framework that combines biologically plausible learning with mathematical modelling for interpretable and efficient lesion detection.

NeuroMorphFusion integrates a lightweight ResNet18 backbone, a Spiking Neural Network (SNN) component to capture temporal dynamics, and a morphological attention mechanism that emphasizes structure-relevant regions in CT scans. The architecture employs a semi-supervised reinforcement learning strategy, where pseudo-label accuracy and the overlap between Grad-CAM visualizations and expert annotations define the reward, ensuring explainable updates under limited labelled data. Additionally, a genetic algorithm (GA) optimizes hyperparameters—learning rate, dropout rate, spiking time steps, and attention dimensionality – for domain generalization and reduced memory use. The optimization population is restricted to 20 individuals over 30 generations, converging within eight minutes on a Jetson Nano.

A multi-objective optimization scheme balances lesion detection sensitivity, computational latency and explainability. Integrated SHAP and Grad-CAM visualizations enhance interpretability. Experimental evaluation on the IQ-OTHNCCD lung cancer CT dataset demonstrates that NeuroMorphFusion achieves 98.18% classification accuracy, outperforming VGG16, SqueezeNet, MobileNetV3, and ResNet18 in both transparency and efficiency.

NeuroMorphFusion effectively unites neuro-biological inspiration, mathematical interpretability, and edge-efficient computation for IoMT-based medical imaging. Its superior accuracy, explainability, and low-latency optimization highlight its potential for real-world clinical integration and scalable IoMT deployment.

Figures