Publicação
Skin Lesion Assessment based on Plenoptic Images for Melanoma Classification
| Resumo: | Over the years, image processing algorithms have achieved many advancements in the medical imaging area, namely in skin lesion detection and classification. Still, skin cancer has maintained its position at the top of the most common cancers all over the world. Early detection of suspicious pigmented skin lesions has a determinant role in clinical prognosis. Among them, melanoma, a malignant type of skin lesions, is the one that causes the most deaths. Several research works have moved forward the methodology and tools employed by expert dermatology clinicians. Currently, most experts employ a dermatoscope in naked eye examination. However, in recent years, some public datasets of dermoscopy images have emerged, enabling researchers to develop, validate, and assess new computer-aided methods. Such methods include: pre-processing algorithms, aimed at removing artefacts and applying transformations necessary for the following algorithms; segmentation methods, that aim at identifying and separating healthy skin from the lesion region; and classification or recognition methods, which aim at detecting key lesion characteristics and even devise the lesion type. However, none of these methods provide sufficient robustness for widespread usage. In the pursuit for further advancements in this field, this thesis addresses and improves current segmentation and classification algorithms, provides a new evaluation tool for dermatology experts and researchers (by introducing a light-field dataset of skin lesion images to the field), and proposes several approaches based on algorithms capable of differentiating melanoma from non-melanoma images using 2D and 3D features. Tackling the challenges in the literature, this thesis first proposes two segmentation approaches, while also performing extensive comparisons with other works, across multiple datasets and performance metrics. From this endeavour, evidence that segmentation-detail can contribute for melanoma discrimination is presented. Using the Light-field Image Dataset of Skin Lesions (SKINL2), with images collected at the Department of Dermatology of Centro Hospitalar de Leiria (Portugal), several methods are presented as the key contributions of this thesis. First, the acquired skin surface depth is explored, confirming that the use of depth data presents relevant information for melanoma classification (data not present in 2D colour images). Then, further steps are taken to exploit both colour and depth information under a joint process, whilst maintaining the capability of showing the depth contribution to the classification performance. In any of these steps, proposed approaches provide results superior the current state-of-the-art, when applied to the SKINL2 dataset. |
|---|---|
| Autores principais: | Pereira, Pedro Miguel Marques |
| Assunto: | Medical Image Analysis Dermoscopy Skin Lesions Melanoma Medical dataset Image Segmentation Feature Extraction Image Classification Análise de Imagens Médicas Dermatoscopia Lesões Cutâneas Melanoma Dataset Médico Segmentação de Imagens Extração de Characterísticas Classificação de Imagens |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso embargado |
| Instituição associada: | Universidade de Coimbra |
| Idioma: | inglês |
| Origem: | Estudo Geral - Universidade de Coimbra |
| Resumo: | Over the years, image processing algorithms have achieved many advancements in the medical imaging area, namely in skin lesion detection and classification. Still, skin cancer has maintained its position at the top of the most common cancers all over the world. Early detection of suspicious pigmented skin lesions has a determinant role in clinical prognosis. Among them, melanoma, a malignant type of skin lesions, is the one that causes the most deaths. Several research works have moved forward the methodology and tools employed by expert dermatology clinicians. Currently, most experts employ a dermatoscope in naked eye examination. However, in recent years, some public datasets of dermoscopy images have emerged, enabling researchers to develop, validate, and assess new computer-aided methods. Such methods include: pre-processing algorithms, aimed at removing artefacts and applying transformations necessary for the following algorithms; segmentation methods, that aim at identifying and separating healthy skin from the lesion region; and classification or recognition methods, which aim at detecting key lesion characteristics and even devise the lesion type. However, none of these methods provide sufficient robustness for widespread usage. In the pursuit for further advancements in this field, this thesis addresses and improves current segmentation and classification algorithms, provides a new evaluation tool for dermatology experts and researchers (by introducing a light-field dataset of skin lesion images to the field), and proposes several approaches based on algorithms capable of differentiating melanoma from non-melanoma images using 2D and 3D features. Tackling the challenges in the literature, this thesis first proposes two segmentation approaches, while also performing extensive comparisons with other works, across multiple datasets and performance metrics. From this endeavour, evidence that segmentation-detail can contribute for melanoma discrimination is presented. Using the Light-field Image Dataset of Skin Lesions (SKINL2), with images collected at the Department of Dermatology of Centro Hospitalar de Leiria (Portugal), several methods are presented as the key contributions of this thesis. First, the acquired skin surface depth is explored, confirming that the use of depth data presents relevant information for melanoma classification (data not present in 2D colour images). Then, further steps are taken to exploit both colour and depth information under a joint process, whilst maintaining the capability of showing the depth contribution to the classification performance. In any of these steps, proposed approaches provide results superior the current state-of-the-art, when applied to the SKINL2 dataset. |
|---|