Back to search

Institutes outside Greece

Hyperspectral imaging and analysis for the diagnosis of pigmented skin lesions during gross pathology

Abstract

dc:description

The dissertation entitled “Hyperspectral Imaging and Analysis for the Diagnosis of Pigmented Skin Lesions during Gross Pathology” consists of 8 chapters.Chapter 1 “Introduction” presents the background, rationale, and goal of the present work. The process and obstacles of gross pathology, the characteristics of skin tissue and the potential of Hyperspectral Imaging (HSI) in diagnosis are presented. The problem statement, research questions, significance and originality are described. This work aims to investigate the particularities of the diagnosis of Pigmented Skin Lesions (PSL) during gross pathology and to propose a framework for tumor margin detection. The goal is to determine the appropriateness of HSI data, tissue condition, preprocessing schemes, and segmentation models for the task of tumor segmentation.Chapter 2 “Literature Review” investigates in detail previous applications of HSI on skin tissue analysis under the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines. Trends in the design of acquisition systems, preprocessing and decision-making are identified. The current research gaps are presented. Chapter 3 “Materials” introduces the data acquisition systems, a Multispectral Imaging (MSI) and a HSI system, and two respective original datasets of PSL, including the data collection and labeling protocols. The information captured by each acquisition system and examples of data samples are explained. A custom HSI acquisition system with high spectral resolution and a wide field of view was designed and calibrated. Evaluation metrics related to further experimentation are presented.Chapter 4 “Effectiveness of Hyperspectral Imaging for Tumor Classification” investigates the effectiveness of HSI data against RGB data. HSI signatures were reconstructed from both MSI and RGB data. The classification of signatures as tumor or not was validated and evaluated using machine learning. A framework for visualization of the results was proposed. Reflectance and texture features from HSI were superior to RGB for classification as either tumor or non-neoplastic growth. Chapter 5 “Influence of Formalin-Fixing on Chromophore Concentrations” investigates the two available types of tissue during gross pathology, namely untreated and formalin-fixed tissue. When anti-infective protocols are in place, i.e., during a pandemic, only formalin-fixed samples are accessible. The optical properties of tissue depend on the concentrations of skin chromophores. Maps of relative concentrations of melanin and hemoglobin were generated from HSI image pairs of tissue, before and after formalin-fixing. Melanin content is mostly retained after formalin-fixing, but hemoglobin content is altered. While formalin-fixed tissue might be sufficient for the analysis of melanocytic lesions, untreated tissue should be used for HSI analysis of PSL. Chapter 6 “Effect of Dimension Reduction on Tumor Margin Detection” investigates the influence of dimension reduction on tumor segmentation. HSI spectral signatures are hundred-point long vectors, which introduce redundancies and affect complexity. No guidelines are available for dimension reduction of HSI from gross skin tissue. Eleven dimension reduction methods were used to compress spectral signatures and train a classifier. A new method based on clustering and spectral similarity was proposed. The method, training scope and number of retained dimensions affected segmentation performance. Standard methods like Principal Component Analysis do not necessarily perform well. The appropriateness of dimension reduction should be confirmed with the target data before further experimentation.Chapter 7 “Implementation and Feasibility of HSI-based Tumor Margin Detection” investigates the feasibility of non-invasive tumor segmentation for PSL during gross pathology. Segmentation is a new field for HSI analysis in medicine. Two frameworks were proposed, one for pixel-wise analysis and one for patch-based analysis. Both machine learning and deep learning models were investigated as core components of the frameworks. These models were based on abundance maps, clustering, Support Vector Machines and three-dimensional convolutional networks. The pixel-wise framework showed the best performance in terms of Jaccard coefficient but suffered from false positives on tissue edges and cannot be easily scaled for a larger dataset. A custom model based on Xception architecture and 3D convolution performed similarly to state-of-the-art classification models. It managed to detect curvy tumor segments, while it has potential for improvement by adjusting the structure depth and kernel for spectral convolution. The patch-based framework managed to segment dense tumor masks, indicating that including both spatial and spectral information is beneficial. Chapter 8 “Conclusions” culminates this work by summarizing the findings of the study, highlighting the limitations, and providing directions for future research. A custom HSI system and a framework for tumor margin detection of PSL during gross pathology were proposed and achieved promising results. Insights about the system and model-building provide valuable knowledge about the handling of HSI data for computer-assisted diagnosis in gross pathology of PSL. The frameworks proposed in this work can be optimized to digitize and automate gross pathology, or even be extended for the early diagnosis of in-situ skin cancers. The focus of system evaluation should shift towards the explainability and robustness of the decision-making process, to use HSI for tumor margin detection in clinical practice.

Degree

thesis:*
Grantor dc:publisher
Institutes outside Greece
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Aloupogianni, Eleni
  • Αλουπογιάννη, Ελένη

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61933

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Aloupogianni, Eleni; Αλουπογιάννη, Ελένη. Hyperspectral imaging and analysis for the diagnosis of pigmented skin lesions during gross pathology. Institutes outside Greece, 2022. http://hdl.handle.net/10442/hedi/61933