Publicação
New algorithms for pigmentation abnormalities detection in human skin
| Resumo: | Detecting and localising regions of hyperpigmentation on facial skin is a challenging task in dermatology. These lesions are often diffuse and irregular, making them difficult to detect and track over time. Accurate localisation and classification of these regions are crucial for dermatological assessments and for companies like PhD Trials®, which evaluate the efficacy of topical products. The ability to extract precise detection metrics can significantly enhance the evaluation of treatment effectiveness, providing deeper insights into skin response. This dissertation presents a deep learning approach for developing a robust automatic object detection system for brownish hyperpigmentation on facial skin. Several pre- trained models, including YOLO and Faster R-CNN, were explored, along with different frameworks such as MATLAB and Detectron2. The proposed solution employs transfer learning to fine-tune these models with a custom dataset, optimising them for detecting and distinguishing lentigines from nevi. The research further extends the utility of these models by integrating segmentation using Otsu’s method, which enabled the delineation of the regions of interest and the extraction of monitoring metrics for each detected area, offering a more comprehensive evaluation of skin health. The results show that this solution outperforms PhD Trials®’ current software, providing superior detection and enhanced usability. The tool facilitates more reliable analyses in studies involving topical treatments, which are valuable for both clinical and cosmetic evaluations. Its ability to work with high-resolution camera images, not just dermoscopic ones, makes it suitable for broader clinical use, including telemedicine. This advancement represents a significant contribution to dermatological research and the development of AI-driven skin analysis tools. |
|---|---|
| Autores principais: | Silva, Liliana Cristina dos Santos |
| Assunto: | Hyperpigmentation Deep Learning Transfer Learning YOLO Faster R-CNN Image Segmentation |
| Ano: | 2024 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Detecting and localising regions of hyperpigmentation on facial skin is a challenging task in dermatology. These lesions are often diffuse and irregular, making them difficult to detect and track over time. Accurate localisation and classification of these regions are crucial for dermatological assessments and for companies like PhD Trials®, which evaluate the efficacy of topical products. The ability to extract precise detection metrics can significantly enhance the evaluation of treatment effectiveness, providing deeper insights into skin response. This dissertation presents a deep learning approach for developing a robust automatic object detection system for brownish hyperpigmentation on facial skin. Several pre- trained models, including YOLO and Faster R-CNN, were explored, along with different frameworks such as MATLAB and Detectron2. The proposed solution employs transfer learning to fine-tune these models with a custom dataset, optimising them for detecting and distinguishing lentigines from nevi. The research further extends the utility of these models by integrating segmentation using Otsu’s method, which enabled the delineation of the regions of interest and the extraction of monitoring metrics for each detected area, offering a more comprehensive evaluation of skin health. The results show that this solution outperforms PhD Trials®’ current software, providing superior detection and enhanced usability. The tool facilitates more reliable analyses in studies involving topical treatments, which are valuable for both clinical and cosmetic evaluations. Its ability to work with high-resolution camera images, not just dermoscopic ones, makes it suitable for broader clinical use, including telemedicine. This advancement represents a significant contribution to dermatological research and the development of AI-driven skin analysis tools. |
|---|