Tommasino, Cristian (2023) Distilled Nuclei Segmentation For Graph-Based Whole Slide Images Analysis and Retrieval. [Tesi di dottorato]

Anteprima

Testo Tommasino_Cristian_36.pdf Download (29MB) | Anteprima

Abstract

The integration of technological advancements, particularly in data processing and machine learning, has profoundly impacted the trajectory of medicine and healthcare evolution. Pathology, an essential pillar of medical diagnosis, has not been exempt from this technological metamorphosis. The conventional methodology of microscope-based histological tissue analysis has undergone a significant transition, culminating in the birth of digital pathology. This digitization has not only enhanced the operational efficiency of pathologists but has also facilitated the inception of an interdisciplinary domain termed computational pathology. This discipline employs computational methodologies to analyze and model histopathological imagery meticulously. The overarching objective of computational pathology is to architect a robust digital diagnostic infrastructure functioning as a Computer-Aided Diagnosis (CAD) system. This paradigmatic shift harbors the potential to radically transform the methodologies employed in the diagnosis and therapeutic interventions of diseases, with a particular emphasis on oncological conditions. The salience of computational pathology is accentuated by its prospective capacity to engender transformative alterations in the diagnostic and therapeutic modalities of oncological diseases. Given the accelerated advancements in deep learning paradigms and computer vision algorithms, in conjunction with the seamless integration of data derived from digital pathology, computational pathology is poised at the precipice of a vast paradigmatic evolution. In this regard, we present two primary contributions: the first focuses on enhancing Picture Archiving and Communication Systems (PACSs), and the second aims to reduce the time complexity analysis of Whole Slide Images (WSIs). Our first contribution concerns employing deep features in a Content-Based Image Retrieval (CBIR) system that employs deep features to improve retrieval in PACSs. This contrasts with prevalent methods typically using metadata to retrieve WSIs or their segments. Our approach uses visual features to improve retrieval, enabling query by example. Our second contribution encompasses two key improvements: one concerns nuclei instance segmentation and classification, and the other focuses on cell-graph representation and classification. Specifically, we introduce Fast-HoVerNet, a distilled version of HoVerNet that is one of the most used networks for nuclei instance segmentation and classification, which offers less inference time while maintaining comparable segmentation and classification capabilities. Employing Fast-HoVerNet, we introduce a novel method to extract cell features in cell-graph representation, which is faster than existing ones, and unifies the process using a single network for both cell detection and representation. Finally, we prove the effectiveness of our cell graph representation using it for cancer subtype classification. Experimental results demonstrate that our approach achieves results comparable to the current state-of-the-art ones but with reduced time complexity.

Downloads

Actions (login required)

Modifica documento