A sustainable neuromorphic framework for disease diagnosis using digital medical imaging

Background and objective:

In the diagnosis of medical images, neural network classifications can support rapid diagnosis together with existing imaging methods. Although current state-of-the-art deep learning methods can contribute to this image recognition, the aim of the present study is to develop a general classification framework with brain-inspired neural networks. Following this intention, spiking neural network models, also known as third-generation models, are included here to capitalize on their sparse characteristics and capacity to significantly decrease energy consumption. Inspired by the recent development of neuromorphic hardware, a sustainable neural network framework is proposed, leading to an energy reduction down to a thousandth compared to the current state-of-the-art second-generation counterpart of artificial neural networks. Making use of sparse signal transmissions as in the human brain, a neuromorphic algorithm for imaging diagnostics is introduced.

Methods:

A novel, sustainable, brain-inspired spiking neural network is proposed to perform the multi-class classification of digital medical images. The framework comprises branched and densely connected layers described by a Leaky-Integrate and Fire (LIF) neuron model. Backpropagation of discontinuous spiking activations in the forward pass is achieved by surrogate gradients, in this case, fast sigmoid. The data for the spiking neural network is encoded into binary spikes with a latency encoding strategy. The proposed model is evaluated on a publicly available dataset of digital X-rays of chest and compared with an equivalent classical neural network. The models are trained using enhanced and pre-processed X-ray images and are evaluated based on classification metrics.

Results:

The proposed neuromorphic framework had an extremely high classification accuracy of 99.22$\text{\%}$ on an unseen test set, together with high precision and recall figures. The framework achieves this accuracy, all the while consuming 1000 times less electrical power than classical neural network architectures.

Conclusion:

Though there is a loss of information due to encoding, the proposed neuromorphic framework has achieved accuracy close to its second-generation counterpart. Therefore, the benefit of the proposed framework is the high accuracy of classification while consuming a thousandth of the power, enabling a sustainable and accessible add-on for the available diagnostic tools, such as medical imaging equipment, to achieve rapid diagnosis.