The brain encodes information into complex spatiotemporal signals evoked in its building cells. To unveil the details of this process, one has to devise robust methods to both register and alter brain cell signals without compromising either the cellular functions or the original brain network topology. Optogenetic techniques are by far the most promising toward this goal. They combine genetic intervention with two-photon stimulation and two-photon microscopy, mostly complemented by mini-mally to non-invasive procedures. In particular, target brain cell populations can be genetically modified so that they express a specific species of fluorescent molecules. These are classified into sensors and actuators, depending on their behavior. The intensity of light emission in sensors is ruled by the concentration of the chemical species they are sensitive to (e. g. genetically encoded calcium indicators with re-spect to Ca2+). Conversely, actuators act as neuromodulators if stimulated with controlled light, as they are able to modify the cellular permeability (e. g. Opsins)[1]. Recent studies have demonstrated that it is possible to activate actuators with a light beam which negligibly interferes with the imaging one [2, 3], thus allowing an efficient, combined employment of the two procedures. These new all-optical approaches enable the formulation of a novel class of groundbreaking experiments which would eventually test causal hypotheses about the neural code, as it is possi-ble to test the consequences of writing or erasing pieces of information in brain cell activities. [4, 2, 3]. For the systematic storage and analysis of the many collected images recording the brain cell activity, experimenters have to face many non-trivial computational challenges which cannot be tackled by employing conventional methods. Neural imaging has eventually entered the big data era, as it has already happened for many other fields of study. Disentangling the relevant information contained in the collected data from noise and background can be unaffordable without employing high-performance computing algorithms. As the final project for this Master’s thesis, I have developed a parallel 3D clustering algorithm to perform the spatiotemporal segmentation of the active brain cell regions in a series of black and white images recording a signal of interest. Its Python implementation—also boosted by a library written in C—has proven to be reasonably fast and to scale well according to both the strong and the weak scaling paradigms. I realized the present work while working as a postdoctoral researcher in the Center for Neuroscience and Cognitive Systems of the Italian Institute of Technology (IIT), Rovereto (Italy) lead by Dr. Stefano Panzeri. The raw data has been collected at the Optical Approaches to Brain Function Lab lead by Dr. Tommaso Fellin at the IIT headquarters, Genoa (Italy). In compliance to a non-disclosure agreement between these two labs and the Scientific Board of the SISSA-ICTP Master in High-Performance Computing, no details about either the raw data or their processing methods will be unveiled in this thesis, with the exception of a series of computational benchmarks of the developed clustering algorithm and a general overview about its implementation, collected in Chapter 1. This policy has been adopted in order to preserve the confidential status of the ongoing project this work is part of.
A parallel clustering algorithm for image segmentation(2018 Dec 20).
A parallel clustering algorithm for image segmentation
-
2018-12-20
Abstract
The brain encodes information into complex spatiotemporal signals evoked in its building cells. To unveil the details of this process, one has to devise robust methods to both register and alter brain cell signals without compromising either the cellular functions or the original brain network topology. Optogenetic techniques are by far the most promising toward this goal. They combine genetic intervention with two-photon stimulation and two-photon microscopy, mostly complemented by mini-mally to non-invasive procedures. In particular, target brain cell populations can be genetically modified so that they express a specific species of fluorescent molecules. These are classified into sensors and actuators, depending on their behavior. The intensity of light emission in sensors is ruled by the concentration of the chemical species they are sensitive to (e. g. genetically encoded calcium indicators with re-spect to Ca2+). Conversely, actuators act as neuromodulators if stimulated with controlled light, as they are able to modify the cellular permeability (e. g. Opsins)[1]. Recent studies have demonstrated that it is possible to activate actuators with a light beam which negligibly interferes with the imaging one [2, 3], thus allowing an efficient, combined employment of the two procedures. These new all-optical approaches enable the formulation of a novel class of groundbreaking experiments which would eventually test causal hypotheses about the neural code, as it is possi-ble to test the consequences of writing or erasing pieces of information in brain cell activities. [4, 2, 3]. For the systematic storage and analysis of the many collected images recording the brain cell activity, experimenters have to face many non-trivial computational challenges which cannot be tackled by employing conventional methods. Neural imaging has eventually entered the big data era, as it has already happened for many other fields of study. Disentangling the relevant information contained in the collected data from noise and background can be unaffordable without employing high-performance computing algorithms. As the final project for this Master’s thesis, I have developed a parallel 3D clustering algorithm to perform the spatiotemporal segmentation of the active brain cell regions in a series of black and white images recording a signal of interest. Its Python implementation—also boosted by a library written in C—has proven to be reasonably fast and to scale well according to both the strong and the weak scaling paradigms. I realized the present work while working as a postdoctoral researcher in the Center for Neuroscience and Cognitive Systems of the Italian Institute of Technology (IIT), Rovereto (Italy) lead by Dr. Stefano Panzeri. The raw data has been collected at the Optical Approaches to Brain Function Lab lead by Dr. Tommaso Fellin at the IIT headquarters, Genoa (Italy). In compliance to a non-disclosure agreement between these two labs and the Scientific Board of the SISSA-ICTP Master in High-Performance Computing, no details about either the raw data or their processing methods will be unveiled in this thesis, with the exception of a series of computational benchmarks of the developed clustering algorithm and a general overview about its implementation, collected in Chapter 1. This policy has been adopted in order to preserve the confidential status of the ongoing project this work is part of.File | Dimensione | Formato | |
---|---|---|---|
Timoteo Colnaghi.pdf
accesso aperto
Descrizione: MHPC Thesis
Tipologia:
Tesi
Licenza:
Non specificato
Dimensione
2.01 MB
Formato
Adobe PDF
|
2.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.