Nuclear medical imaging visualizes the distribution of injected radioactive tracers to gain insight on physiological processes in-vivo. Clinical scanners currently employ gamma-ray detectors comprising pixelated scintillator crystals coupled with silicon photomultipliers (SiPM). However, detectors based on monolithic scintillators are gaining interest, as they offer higher spatial resolution, but require sophisticated position reconstruction algorithms. Machine learning (ML) methods, and in particular artificial neural networks, proved to be effective in this task. The demand for processing larger volumes of data, along with the need to reduce power consumption and latency, has led to embedding ML near the photosensors. The ANNA ASIC (Analog Neural Network Application-Specific Integrated Circuit), object of experimental characterization in this thesis, was designed to address these demands by integrating an artificial neural network on-chip. Fabricated in a 0.35 µm CMOS process node and occupying a 24 mm² area, it incorporates 64 input neurons, two hidden layers of 20 fully connected neurons each, and two output neurons. This chip is intended to process signals generated by a matrix of 8x8 SiPMs and reconstruct the 2D position (X and Y coordinates) of interaction of gamma rays within a monolithic scintillator. Multiply-and-accumulate operations, required for inference, are performed in the analog domain, thus minimizing the need for analog-to-digital conversion and reducing latency and power consumption. The network weights are implemented by banks of programmable capacitors, and the activation functions are intrinsically applied by charge integrators. During this thesis work, a dedicated test setup was developed to enable characterization of a first ANNA ASIC prototype. An initial debugging phase was instrumental in identifying circuit non-idealities and accurately modelling the physical chip behavior, optimizing off-chip network training. Subsequently, the ASIC performance as neural network has been assessed. Measurements yielded promising results in terms of precision and energy efficiency, and provided valuable insights for future chip design improvements.
La medicina nucleare utilizza marcatori radioattivi per visualizzare processi fisiologici in vivo. Attualmente gli scanner clinici utilizzano rivelatori di raggi gamma che comprendono cristalli scintillatori pixelati accoppiati a fotomoltiplicatori al silicio (SiPM). Tuttavia, i rivelatori basati su scintillatori monolitici stanno guadagnando interesse, in quanto offrono una maggiore risoluzione spaziale, ma richiedono sofisticati algoritmi di ricostruzione della posizione. I metodi di machine learning (ML), e in particolare le reti neurali artificiali, si sono dimostrati efficaci in questo compito. La crescente necessità di elaborare grandi volumi di dati, insieme all'esigenza di ridurre il consumo energetico e la latenza, ha portato a incorporare il ML in prossimità dei fotorivelatori. L'ASIC ANNA (Analog Neural Network Application-Specific Integrated Circuit), oggetto di caratterizzazione sperimentale in questa tesi, è stato progettato per rispondere a queste esigenze integrando una rete neurale artificiale su chip. Realizzato in tecnologia CMOS da 0.35 µm e occupando un'area di 24 mm², incorpora 64 neuroni di ingresso, due hidden layer da 20 neuroni ciascuno e due neuroni di uscita. Il chip è pensato per elaborare i segnali generati da una matrice di 8x8 SiPM e ricostruire la posizione 2D di interazione dei raggi gamma con uno scintillatore monolitico. Le operazioni sono eseguite nel dominio analogico, limitando così la necessità di conversione analogico-digitale e riducendo la latenza e il consumo di energia. I pesi della rete sono implementati da banchi capacitivi programmabili e le funzioni di attivazione sono applicate dagli integratori di cariche. Durante questo lavoro di tesi è stato sviluppato un setup sperimentale dedicato alla caratterizzazione di un primo prototipo di ANNA. Una fase iniziale di debug è servita ad identificare le non-idealità circuitali e modellare il comportamento reale del chip. Successivamente sono state valutate le prestazioni dell'ASIC come rete neurale. Le misure hanno dato risultati promettenti in termini di precisione ed efficienza energetica e hanno fornito indicazioni utili per la progettazione di future versioni del chip.
Experimental characterization of an Analog Neural Network ASIC (ANNA) for 2D gamma-ray positioning in Anger Cameras
Palmieri, Giulia
2023/2024
Abstract
Nuclear medical imaging visualizes the distribution of injected radioactive tracers to gain insight on physiological processes in-vivo. Clinical scanners currently employ gamma-ray detectors comprising pixelated scintillator crystals coupled with silicon photomultipliers (SiPM). However, detectors based on monolithic scintillators are gaining interest, as they offer higher spatial resolution, but require sophisticated position reconstruction algorithms. Machine learning (ML) methods, and in particular artificial neural networks, proved to be effective in this task. The demand for processing larger volumes of data, along with the need to reduce power consumption and latency, has led to embedding ML near the photosensors. The ANNA ASIC (Analog Neural Network Application-Specific Integrated Circuit), object of experimental characterization in this thesis, was designed to address these demands by integrating an artificial neural network on-chip. Fabricated in a 0.35 µm CMOS process node and occupying a 24 mm² area, it incorporates 64 input neurons, two hidden layers of 20 fully connected neurons each, and two output neurons. This chip is intended to process signals generated by a matrix of 8x8 SiPMs and reconstruct the 2D position (X and Y coordinates) of interaction of gamma rays within a monolithic scintillator. Multiply-and-accumulate operations, required for inference, are performed in the analog domain, thus minimizing the need for analog-to-digital conversion and reducing latency and power consumption. The network weights are implemented by banks of programmable capacitors, and the activation functions are intrinsically applied by charge integrators. During this thesis work, a dedicated test setup was developed to enable characterization of a first ANNA ASIC prototype. An initial debugging phase was instrumental in identifying circuit non-idealities and accurately modelling the physical chip behavior, optimizing off-chip network training. Subsequently, the ASIC performance as neural network has been assessed. Measurements yielded promising results in terms of precision and energy efficiency, and provided valuable insights for future chip design improvements.File | Dimensione | Formato | |
---|---|---|---|
2024_07_Palmieri.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
44.76 MB
Formato
Adobe PDF
|
44.76 MB | Adobe PDF | Visualizza/Apri |
2024_07_Palmieri_Executive_Summary.pdf
non accessibile
Descrizione: Executive summary
Dimensione
3.16 MB
Formato
Adobe PDF
|
3.16 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/10589/223412