Recent advances in neural electrophysiology allows the development of improved acquisition systems, as well as novel data analysis and denoising techniques. In this regard, DeepInterpolation, a state-of-the-art denoising algorithm, has been shown to effectively reduce statistically independent noise that contaminates electrophysiological data. Novel high-density silicon-based microelectrodes, such as Neuropixels 1.0 and 2.0, have revolutionized the field of electrophysiology due to their ability to record neuronal activity from hundreds of neurons thanks to their high spatial-density and channel count. However, DeepInterpolation is exclusively designed for Neuropixels 1.0 probes, thus lacking the flexibility to interface with other recording instruments. This thesis project investigates the possibility of extending the application of DeepInterpolation to Neuropixels 2.0. In addition, a second goal of this project is to integrate DeepInterpolation into the API of SpikeInterface, which is a Python framework aiming to unify the analysis of extracellular electrophysiology. To achieve this, the U-Net architecture of Deepinterpolation is adjusted to match the (192x2) geometry of Neuropixels 2.0. Then, DeepInterpolation models are trained on a variety of electrophysiology recordings, using either a band-pass or a high-pass filter as a preprocessing step. Subsequently, during the inference phase, DeepInterpolation-based models are applied to the original recordings, thus generating their “DeepInterpolated” versions. To evaluate the effect of Deepinterpolation, spike sorting with pyKilosort is later conducted on the two versions of the same recording -one with DeepInterpolation processing and the other without- followed by a comprehensive quality and performance analysis at the session and the single-unit levels. From the resulting outcomes, it is concluded that the association of high-pass filter and DeepInterpolation results in sorting a higher number of bad units; However, with band-pass filtered recording, the use of DeepInterpolation did not produce any significant improvement of the quality or performance of the sorting process. These results do not necessarily indicate that DeepInterpolation is not associable with Neuropixels 2.0, but simply that the design choices made for this project, for example with respect to the training phase, might need to be revisited in future work to improve Deepinterpolation performance.
I recenti progressi nell'elettrofisiologia neurale hanno portato allo sviluppo di sistemi di acquisizione all’avanguardia, nonché di nuove tecniche di analisi dei dati e di denoising. A questo proposito, DeepInterpolation, un algoritmo di denoising allo stato dell’arte, riduce efficacemente il rumore statisticamente indipendente che travolge i dati elettrofisiologici. Tuttavia, DeepInterpolation è progettato esclusivamente per le sonde Neuropixels 1.0, e quindi manca della flessibilità di interfacciarsi con altri strumenti di registrazione. Questo progetto di tesi indaga la possibilità di estendere l'applicazione DeepInterpolation alle nuove sonde elettrofisiologiche ad alta densità, Neuropixels 2.0. Inoltre, un secondo obiettivo è l’integrazione di DeepInterpolation nel software di SpikeInterface, un progetto open-source per l’analisi di dati di elettrofisiologia, per facilitarne la disseminazione e l’utilizzo. Per raggiungere questi obiettivi, l'architettura U-Net di DeepInterpolation è stata adattata per corrispondere alla geometria (192x2) di Neuropixels 2.0. Quindi, modelli di DeepInterpolation sono stati addestrati su una larga varietà di registrazioni elettrofisiologiche, utilizzando o un filtraggio passa-banda o un passa-alto come pre-processing. Successivamente, durante la fase di inferenza, i modelli addestrati in precedenza vengono applicati alle registrazioni originali, generando così le relative versioni “DeepInterpolated”. Per valutare l’effetto di DeepInterpolation, abbiamo applicato spike sorting tramite pyKilosort sulle due versioni della stessa registrazione, una processata con Deepinterpolation e l’atra senza, seguito da un’analisi completa della qualità delle prestazioni a livello delle sessioni e delle singole unità. I risultati indicano che l'associazione del filtro passa-alto e della DeepInterpolation comporta l'identificazione di un numero maggiore di unità di bassa qualità; Tuttavia, con la registrazione filtrata passa-banda, l'uso di DeepInterpolation non ha prodotto alcun miglioramento significativo della qualità o delle prestazioni dello spike sorting. Questi risultati non indicano necessariamente che DeepInterpolation non sia compatibile a sonde Neuropixels 2.0, ma semplicemente che le scelte progettuali effettuate per questo lavoro, per esempio riguardo alla fase di addestramento dei modelli, andrebbero rivisitate in futuri per migliorare la performance della DeepInterpolation.
Investigation and extension of Deepinterpolation to denoise high-density electrophysiological recordings
SELMAN, JAD
2022/2023
Abstract
Recent advances in neural electrophysiology allows the development of improved acquisition systems, as well as novel data analysis and denoising techniques. In this regard, DeepInterpolation, a state-of-the-art denoising algorithm, has been shown to effectively reduce statistically independent noise that contaminates electrophysiological data. Novel high-density silicon-based microelectrodes, such as Neuropixels 1.0 and 2.0, have revolutionized the field of electrophysiology due to their ability to record neuronal activity from hundreds of neurons thanks to their high spatial-density and channel count. However, DeepInterpolation is exclusively designed for Neuropixels 1.0 probes, thus lacking the flexibility to interface with other recording instruments. This thesis project investigates the possibility of extending the application of DeepInterpolation to Neuropixels 2.0. In addition, a second goal of this project is to integrate DeepInterpolation into the API of SpikeInterface, which is a Python framework aiming to unify the analysis of extracellular electrophysiology. To achieve this, the U-Net architecture of Deepinterpolation is adjusted to match the (192x2) geometry of Neuropixels 2.0. Then, DeepInterpolation models are trained on a variety of electrophysiology recordings, using either a band-pass or a high-pass filter as a preprocessing step. Subsequently, during the inference phase, DeepInterpolation-based models are applied to the original recordings, thus generating their “DeepInterpolated” versions. To evaluate the effect of Deepinterpolation, spike sorting with pyKilosort is later conducted on the two versions of the same recording -one with DeepInterpolation processing and the other without- followed by a comprehensive quality and performance analysis at the session and the single-unit levels. From the resulting outcomes, it is concluded that the association of high-pass filter and DeepInterpolation results in sorting a higher number of bad units; However, with band-pass filtered recording, the use of DeepInterpolation did not produce any significant improvement of the quality or performance of the sorting process. These results do not necessarily indicate that DeepInterpolation is not associable with Neuropixels 2.0, but simply that the design choices made for this project, for example with respect to the training phase, might need to be revisited in future work to improve Deepinterpolation performance.File | Dimensione | Formato | |
---|---|---|---|
[Final] Investigation and Extension of DeepInterpolation to Denoise High-density Electrophysiological Recordings - Jad Selman-.pdf
accessibile in internet per tutti
Descrizione: The Final Official version of the thesis
Dimensione
6 MB
Formato
Adobe PDF
|
6 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/210023