Back to search

University of the Aegean

Neuromorphic photonic architectures for biomedical and conventional image processing

Abstract

dc:description

This doctoral thesis focuses on the development of a high-speed, power-efficient integrated neuromorphic photonic platform for reservoir computing (RC) and the acceleration of convolutional neural networks (CNNs) regarding biomedical and conventional image classification tasks. Neuromorphic architectures consist of two collocated components: neurons, modeled as biologically inspired spiking neurons or simplified artificial neuron models, and synapses, interconnecting the nodes. RC architectures have random, fixed interconnections between their neurons to project input signals into higher-dimensional spaces, enabling classification through a simple readout layer. On the other hand, CNNs are a subset of deep artificial neural networks which specialize in image and signal classification, performing convolutions in one or more layers of their structure. Implementing these architectures in the photonic domain poses challenges. Regarding photonic CNNs, convolutions are performed through discrete analog matrix-to-vector multiplications, which are prone to errors during weight transfer, require extensive electro-optic conversions, and demand multiple chip reconfigurations. In this thesis an alternative is presented based on an optical spectrum slicing (OSS) technique. This OSS approach employs multiple parallel optical filter nodes, each “slicing” a different spectral region of the optical signal, with these lower-bandwidth components photodetected and digitized before fed on a digital readout layer. Through this technique, convolutions occur in the analog temporal domain through the interaction of modulated optical signals with each OSS node’s complex-valued impulse response, with additional nonlinear and averaging operations performed during photodetection. Numerical simulations on a handwritten digit classification task showed that using the OSS preprocessing with 7 microring resonator nodes in front of a single-layer fully connected digital readout improved accuracy up to 5.4% compared to the direct readout processing. This impact of the OSS was also experimentally validated for this task using a reconfigurable photonic chip to implement the OSS nodes, achieving a comparable performance on a noise-present realistic photonic platform. Compared to a respective digital CNN, the OSS neural network (OSS-NN) demonstrated a 30% smaller total power consumption, while maintaining comparable precision. Furthermore, a comparative analysis revealed that the OSS accelerator achieved 7–13 times better power efficiency than state-of-the-art electronic and photonic CNN accelerators, demonstrating also high computational density. On an event-based imaging flow cytometry (IFC) classification task, the OSS-NN improved performance by 0.4%, reducing input digital samples, learnable parameters, and floating-point operations during inference by ~15× for a single and ~9× for a two dense layer readout, while it matched the precision of a GRU-RNN with 22× fewer parameters. Experimental validation confirmed OSS-NN’s performance for this task, resilient to noise and data compression. For a time-stretched IFC signal classification task, the OSS-NN enhanced accuracy by 0.2%, reducing trainable parameters 10× and 13× for a single and a two dense layer readout respectively. Lastly, a spatially distributed RC, comprising OSS nodes within a common recurrent loop, was used to directly process the optical output of a simulated time-stretched IFC system regarding a cell classification task, improving the accuracy of a single fully connected readout by 3%, while it reduced the photodetection and digitization demands by 10× and 5×, respectively.

Degree

thesis:*
Grantor dc:publisher
University of the Aegean
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Tsirigotis, Aris
  • Τσιριγώτης, Άρης

Subjects

dc:subject × 24

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60330

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Tsirigotis, Aris; Τσιριγώτης, Άρης. Neuromorphic photonic architectures for biomedical and conventional image processing. University of the Aegean, 2025. http://hdl.handle.net/10442/hedi/60330