During my PhD research activity I developed novel AI-based algorithms devoted to the analysis of signals and images. In particular, they couple the Discrete Fourier Transform (DFT) and its related properties and analysis techniques (as the phasor approach) with Machine Learning (ML) techniques. These novel approaches have been exploited in two different research fields: the histopathology and the remote sensing. During my first project, in collaboration with the Remote Sensing laboratory of the Milano-Bicocca University (Prof. Colombo, Dr. Cogliati), I developed i-φ-MaLe, an algorithm able to provide reliable estimations of both many biophysical parameters characterizing the vegetation, as the fluorescence quantum yields induced by the sunlight, and fluorescence spectra. In particular, i-φ-MaLe analyzed spectra experimentally acquired by spectrometers at different distance scales (from top of canopy level to 100m from the vegetation). i-φ-MaLe has been validated by exploiting simulations and has been applied to experimental data acquired at seasonal and diurnal timescale on both agricultural cultivations and forests. The results have been compared with field measurements, highlighting a strong compatibility between them. The related article has been published on the peer-reviewed journal “Remote Sensing of Environment” (I.F. 13.85). During my second project I developed SuperµMAPPS. This algorithm analyzes the Second Harmonic Generation signal acquired in dependence on the laser polarization to provide the mean orientation angle and the mean anisotropy characterizing the collagen fibrils. These parameters are affected by pathologies, as tumor growth. The preliminary results, obtained by analyzing images of murine tendons, demonstrate that SuperµMAPPS provides results which are compatible with the literature by exploiting only 6 experimental points per spectrum, reducing the acquisition time by 65%. The values retrieved by SuperµMAPPS during the analysis of tumorous human samples (extracted from the public repository PSHG-TISS) are compatible with the results provided by standard interpolation procedures and, if coupled with clustering techniques, are able to classify biopsies characterized by tumorous regions and healthy ones from the early stage of tumor growth. During the third project, I developed Φ-Norm, algorithm able to analyze images of entire biopsies acquired by whole slide scanners and normalize the color expressed by staning procedures performed in different laboratories. The comparison with other color normalization techniques demonstrates that this novel approach is the most suitable to be coupled with artificial intelligence methods devoted to the automated semantic segmentation of biological structures in order to improve the accuracy of these pipelines. In summary, these studies demonstrate the possibility to boost the geometrical and mathematical properties of the DFT by exploiting supervised machine learning techniques. The application of the three proposed tools, i-φ-MaLe, SuperµMAPPS and Φ-Norm, in different research fields proves the outstanding versatility of the novel pipeline I conceived during my PhD project. In particular, these studies pave the way to new research fields concerning the DFT and image analysis by breathing new life into the standard phasor approach.
Durante il Corso di Dottorato ho sviluppato nuovi algoritmi di intelligenza artificiale che analizzano segnali e immagini sfruttando alcune proprietà della Trasformata di Fourier Discreta (DFT) e dei metodi di analisi che la sfruttano, in particolare l’approccio dei fasori, e accoppiandole con tecniche di Machine Learning (ML). Questi algoritmi sono stati applicati a progetti relativi sia al campo biofisico e medico che a quello del telerilevamento. Nel primo progetto, in collaborazione con il Dipartimento di Scienze Ambientali (Prof. Colombo, Dr. Cogliati), ho sviluppato l’algoritmo i-φ-MaLe, che stima simultaneamente diversi parametri biofisici caratteristici della vegetazione, la resa quantica di fluorescenza della clorofilla indotta dalla luce solare e gli spettri di fluorescenza a livello di chioma e di fotosistema a partire dagli spettri di radianza riflessa, acquisiti mediante sensori situati fino a 100m dal suolo. Il metodo, validato tramite simulazioni, è stato applicato a dati sperimentali acquisiti su base giornaliera e annuale in campi agricoli e foreste. I risultati sono stati validati tramite dati acquisiti sul campo. L’articolo relativo al progetto è stato pubblicato sulla rivista “Remote Sensing of Environment” (I.F. 13.85). Nel secondo progetto ho sviluppato SuperµMAPPS, algoritmo che analizza il segnale di generazione di seconda armonica in funzione della polarizzazione del laser incidente per stimare l’angolo di orientazione e l’anisotropia media delle fibrille di collagene, parametri influenzati dalla presenza di patologie, quali lo sviluppo tumorale. I primi risultati, ottenuti analizzando immagini di tendine murino, indicano che l’approccio fornisce risultati comparabili con quelli riportati in letteratura sfruttando solo 6 punti sperimentali, così velocizzando la fase di acquisizione del 65%. I valori forniti da SuperµMAPPS durante l’analisi di immagini relative a diversi tipi di tumore umani (presenti nella libreria PSHG-TISS) sono compatibili con quelli ottenuti da procedure standard di interpolazione e, accoppiati con metodi di clustering, riconoscono le biopsie caratterizzate da aree tumorali rispetto a quelle in cui è presente solo tessuto sano fin dai primi stadi di sviluppo della malattia. Nel terzo progetto, ho sviluppato Φ-Norm, algoritmo che analizza le immagini di biopsie acquisite tramite scanner digitali e uniforma le colorazioni ottenute tramite procedure di staining effettuate in diversi laboratori. Il confronto con altre tecniche di normalizzazione del colore ha dimostrato che il nuovo approccio è il migliore per essere accoppiato a tecniche di intelligenza artificiale volte alla segmentazione di strutture biologiche all’interno di immagini al fine di migliorare l’accuratezza di questi metodi. Questi metodi dimostrano la possibilità di sfruttare in modo innovativo le proprietà geometriche e matematiche della DFT attraverso l’utilizzo di tecniche di intelligenza artificiale supervisionate. L’applicazione di questi algoritmi (i-φ-MaLe, SuperµMAPPS and Φ-Norm) in differenti campi di ricerca dimostra la grande versatilità dell’approccio che ho sviluppato durante il Corso di Dottorato. In particolare, questi studi aprono la strada a nuovi campi di ricerca riguardanti la relazione tra l’applicazione della DFT e l’analisi di immagini.
NOVEL HYBRID AI-PHASOR BASED TOOLS FOR REMOTE SENSING AND HISTOPATHOLOGICAL APPLICATIONS
SCODELLARO, RICCARDO
2023
Abstract
During my PhD research activity I developed novel AI-based algorithms devoted to the analysis of signals and images. In particular, they couple the Discrete Fourier Transform (DFT) and its related properties and analysis techniques (as the phasor approach) with Machine Learning (ML) techniques. These novel approaches have been exploited in two different research fields: the histopathology and the remote sensing. During my first project, in collaboration with the Remote Sensing laboratory of the Milano-Bicocca University (Prof. Colombo, Dr. Cogliati), I developed i-φ-MaLe, an algorithm able to provide reliable estimations of both many biophysical parameters characterizing the vegetation, as the fluorescence quantum yields induced by the sunlight, and fluorescence spectra. In particular, i-φ-MaLe analyzed spectra experimentally acquired by spectrometers at different distance scales (from top of canopy level to 100m from the vegetation). i-φ-MaLe has been validated by exploiting simulations and has been applied to experimental data acquired at seasonal and diurnal timescale on both agricultural cultivations and forests. The results have been compared with field measurements, highlighting a strong compatibility between them. The related article has been published on the peer-reviewed journal “Remote Sensing of Environment” (I.F. 13.85). During my second project I developed SuperµMAPPS. This algorithm analyzes the Second Harmonic Generation signal acquired in dependence on the laser polarization to provide the mean orientation angle and the mean anisotropy characterizing the collagen fibrils. These parameters are affected by pathologies, as tumor growth. The preliminary results, obtained by analyzing images of murine tendons, demonstrate that SuperµMAPPS provides results which are compatible with the literature by exploiting only 6 experimental points per spectrum, reducing the acquisition time by 65%. The values retrieved by SuperµMAPPS during the analysis of tumorous human samples (extracted from the public repository PSHG-TISS) are compatible with the results provided by standard interpolation procedures and, if coupled with clustering techniques, are able to classify biopsies characterized by tumorous regions and healthy ones from the early stage of tumor growth. During the third project, I developed Φ-Norm, algorithm able to analyze images of entire biopsies acquired by whole slide scanners and normalize the color expressed by staning procedures performed in different laboratories. The comparison with other color normalization techniques demonstrates that this novel approach is the most suitable to be coupled with artificial intelligence methods devoted to the automated semantic segmentation of biological structures in order to improve the accuracy of these pipelines. In summary, these studies demonstrate the possibility to boost the geometrical and mathematical properties of the DFT by exploiting supervised machine learning techniques. The application of the three proposed tools, i-φ-MaLe, SuperµMAPPS and Φ-Norm, in different research fields proves the outstanding versatility of the novel pipeline I conceived during my PhD project. In particular, these studies pave the way to new research fields concerning the DFT and image analysis by breathing new life into the standard phasor approach.File | Dimensione | Formato | |
---|---|---|---|
phd_unimib_769386.pdf
embargo fino al 10/05/2026
Dimensione
6.7 MB
Formato
Adobe PDF
|
6.7 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/76342
URN:NBN:IT:UNIMIB-76342