Publicação
Designing quality metrics for colonoscopy video using explainable AI techniques
| Resumo: | Deep Learning has emerged as a powerful tool in medical image analysis, attracting significant attention due to its potential to improve diagnostic accuracy. In the context of Colorectal Cancer - the third most commonly diagnosed cancer worldwide - early detection via colonoscopy substantially increases the odds of survival. While several approaches with Artificial Intelligence have developed an automated support system for polyp detection and characterization, they are often challenged by variations in colonoscopy image quality. With this in mind, the primary focus of this work is to leverage explainable Artificial Intelligence to understand model decision-making and enhance model reliability, especially for clinical applications. By comparing explainability maps generated by Artificial Intelligence models with ground truth segmentation of polyps, the study assesses whether the model’s focus aligns with clinically relevant regions. The underlying hypothesis is that accurate predictions are more likely when the model’s attention is directed towards the polyp rather than surrounding noise. This dissertation has four core objectives: (1) to explore different explainable Artificial Inteligence techniques, (2) to calculate metrics that evaluate the spatial overlap between explainability maps and segmentation masks, thereby identifying high- and low-quality images automatically; (3) to correlate the new metrics with the model classification performance, and (4) to investigate the potential use of the new metrics to train a model for estimating the adequacy of images for further analysis. Ultimately, the results of this research aim to enhance the reliability and transparency of Artificial Intelligence models in Colorectal Cancer detection among healthcare professionals and improve clinical outcomes. |
|---|---|
| Autores principais: | Martins, Andreia |
| Assunto: | Artificial Intelligence Explainability Colonoscopy videos Colorectal polyp classification Classificação de pólipos colorretais Explicabilidade Inteligência Artificial Vídeos de colonoscopia Engenharia e Tecnologia::Engenharia Médica |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Deep Learning has emerged as a powerful tool in medical image analysis, attracting significant attention due to its potential to improve diagnostic accuracy. In the context of Colorectal Cancer - the third most commonly diagnosed cancer worldwide - early detection via colonoscopy substantially increases the odds of survival. While several approaches with Artificial Intelligence have developed an automated support system for polyp detection and characterization, they are often challenged by variations in colonoscopy image quality. With this in mind, the primary focus of this work is to leverage explainable Artificial Intelligence to understand model decision-making and enhance model reliability, especially for clinical applications. By comparing explainability maps generated by Artificial Intelligence models with ground truth segmentation of polyps, the study assesses whether the model’s focus aligns with clinically relevant regions. The underlying hypothesis is that accurate predictions are more likely when the model’s attention is directed towards the polyp rather than surrounding noise. This dissertation has four core objectives: (1) to explore different explainable Artificial Inteligence techniques, (2) to calculate metrics that evaluate the spatial overlap between explainability maps and segmentation masks, thereby identifying high- and low-quality images automatically; (3) to correlate the new metrics with the model classification performance, and (4) to investigate the potential use of the new metrics to train a model for estimating the adequacy of images for further analysis. Ultimately, the results of this research aim to enhance the reliability and transparency of Artificial Intelligence models in Colorectal Cancer detection among healthcare professionals and improve clinical outcomes. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.