Nowadays, the biomedical signal processing and classification and medical image interpretation play an essential role in the detection and diagnosis of several human diseases. The problem of high variability and heterogeneity of information, which is extracted from digital data, can be addressed with signal decomposition and stain separation techniques which can be useful approaches to highlight hidden patterns or rhythms in biological signals and specific cellular structures in histological color images, respectively. This thesis work can be divided into two macro-sections. In the first part (Part I), a novel cascaded RNN model based on long short-term memory (LSTM) blocks is presented with the aim to classify sleep stages automatically. A general workflow based on single-channel EEG signals is developed to enhance the low performance in staging N1 sleep without reducing the performances in the other sleep stages (i.e. Wake, N2, N3 and REM). In the same context, several signal decomposition techniques and time-frequency representations are deployed for the analysis of EEG signals. All extracted features are analyzed by using a novel correlation-based timestep feature selection and finally the selected features are fed to a bidirectional RNN model. In the second part (Part II), a fully automated method named SCAN (Stain Color Adaptive Normalization) is proposed for the separation and normalization of staining in digital pathology. This normalization system allows to standardize digitally, automatically and in a few seconds, the color intensity of a tissue slide with respect to that of a target image, in order to improve the pathologist’s diagnosis and increase the accuracy of computer-assisted diagnosis (CAD) systems. Multiscale evaluation and multi-tissue comparison are performed for assessing the robustness of the proposed method. In addition, a stain normalization based on a novel mathematical technique, named ICD (Inverse Color Deconvolution) is developed for immunohistochemical (IHC) staining in histopathological images. In conclusion, the proposed techniques achieve satisfactory results compared to state-of-the-art methods in the same research field. The workflow proposed in this thesis work and the developed algorithms can be employed for the analysis and interpretation of other biomedical signals and for digital medical image analysis.

Analysis of Signal Decomposition and Stain Separation methods for biomedical applications / Michielli, Nicola. - (2021).

Analysis of Signal Decomposition and Stain Separation methods for biomedical applications

Michielli, Nicola
2021

Abstract

Nowadays, the biomedical signal processing and classification and medical image interpretation play an essential role in the detection and diagnosis of several human diseases. The problem of high variability and heterogeneity of information, which is extracted from digital data, can be addressed with signal decomposition and stain separation techniques which can be useful approaches to highlight hidden patterns or rhythms in biological signals and specific cellular structures in histological color images, respectively. This thesis work can be divided into two macro-sections. In the first part (Part I), a novel cascaded RNN model based on long short-term memory (LSTM) blocks is presented with the aim to classify sleep stages automatically. A general workflow based on single-channel EEG signals is developed to enhance the low performance in staging N1 sleep without reducing the performances in the other sleep stages (i.e. Wake, N2, N3 and REM). In the same context, several signal decomposition techniques and time-frequency representations are deployed for the analysis of EEG signals. All extracted features are analyzed by using a novel correlation-based timestep feature selection and finally the selected features are fed to a bidirectional RNN model. In the second part (Part II), a fully automated method named SCAN (Stain Color Adaptive Normalization) is proposed for the separation and normalization of staining in digital pathology. This normalization system allows to standardize digitally, automatically and in a few seconds, the color intensity of a tissue slide with respect to that of a target image, in order to improve the pathologist’s diagnosis and increase the accuracy of computer-assisted diagnosis (CAD) systems. Multiscale evaluation and multi-tissue comparison are performed for assessing the robustness of the proposed method. In addition, a stain normalization based on a novel mathematical technique, named ICD (Inverse Color Deconvolution) is developed for immunohistochemical (IHC) staining in histopathological images. In conclusion, the proposed techniques achieve satisfactory results compared to state-of-the-art methods in the same research field. The workflow proposed in this thesis work and the developed algorithms can be employed for the analysis and interpretation of other biomedical signals and for digital medical image analysis.
File in questo prodotto:
File Dimensione Formato  
Thesis_PhD_NM.pdf

Open Access dal 09/04/2022

Descrizione: PhD Thesis
Tipologia: Tesi di dottorato
Licenza: Creative commons
Dimensione 27.71 MB
Formato Adobe PDF
27.71 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2888374
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo