Publicação
Detection, segmentation and classification of polyps in colonoscopies with mask R-CNN
| Resumo: | According to the world’s health data, colorectal cancer is globally the third most prevalent cancer and it’s mainly derived from pre-malignant coloretal polyps. With the goal of decreasing the incidence and the mortality rate associated to this cancer, early diagnose and treatment is the most effective method, achieved through colonoscopy screening. During a colonoscopy, the gastrointestinal tract is visually examined through an endoscope, which is a lighted and flexible tube with a tiny video camera. However, some anomalies may not be detected during a colonoscopy due to human error. With the aim to eliminate this problem, over the last years, many studies have been conducted on computer-aided detection based on deep learning, to reduce the miss rate. Deep learning is an artificial intelligence function that simulates the workings of the human brain in processing data and creating patterns for use in decision making. Deep Learning uses a hierarchical level of artificial neural networks to perform the process of machine learning, enabling transfer learning, where parameters of each layer are changed based on information from the previous layer. Artificial neural networks are built like the human brain, with nodes of neurons linked together like a web. Nowadays, with the incessant advances in information technology and its implications in all the life domains, deep learning algorithms started to emerge as a need for better machine performance, compensating for human’s limited processing capability, preventing human errors, and giving machines some reliable autonomy. This project aimed to develop a model that allows the detection and classification of polyps using Mask R-CNN and evaluate the integration approach with the FUJIFILM Endoscopy Solution. The first key of this dissertation, the Mask R-CNN model, was effectively developed and proven to achieve accurate polyp detection and segmentation. Based on the models that obtain the best metrics during the training phase, it was possible to find a final model that presented an accuracy of 83.33 %, a precision of 93.90 %, a recall of 82.35 %, an F1 Score of 87.75 %, a dice coefficient of 87.51 % and a medium time of 0.22 seconds per prediction. Finally, the integration evaluation was based on two different approaches: a Flask API and Python.Net. Even though the first one was shown to be a good approach with easy integration, the latency in the data transfer process was fatal for what was proposed. The developed methods and implementations proved that, with some improvements, they could successfully create a polyp detection and segmentation system. |
|---|---|
| Autores principais: | Soares, Filipa Ferreira |
| Assunto: | Polyp detection Segmentation Artificial intelligence Deep learning Colonoscopy Deteção de pólipos Segmentação Inteligência artificial Deep learning Colonoscopia Engenharia e Tecnologia::Engenharia Médica |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso restrito |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | According to the world’s health data, colorectal cancer is globally the third most prevalent cancer and it’s mainly derived from pre-malignant coloretal polyps. With the goal of decreasing the incidence and the mortality rate associated to this cancer, early diagnose and treatment is the most effective method, achieved through colonoscopy screening. During a colonoscopy, the gastrointestinal tract is visually examined through an endoscope, which is a lighted and flexible tube with a tiny video camera. However, some anomalies may not be detected during a colonoscopy due to human error. With the aim to eliminate this problem, over the last years, many studies have been conducted on computer-aided detection based on deep learning, to reduce the miss rate. Deep learning is an artificial intelligence function that simulates the workings of the human brain in processing data and creating patterns for use in decision making. Deep Learning uses a hierarchical level of artificial neural networks to perform the process of machine learning, enabling transfer learning, where parameters of each layer are changed based on information from the previous layer. Artificial neural networks are built like the human brain, with nodes of neurons linked together like a web. Nowadays, with the incessant advances in information technology and its implications in all the life domains, deep learning algorithms started to emerge as a need for better machine performance, compensating for human’s limited processing capability, preventing human errors, and giving machines some reliable autonomy. This project aimed to develop a model that allows the detection and classification of polyps using Mask R-CNN and evaluate the integration approach with the FUJIFILM Endoscopy Solution. The first key of this dissertation, the Mask R-CNN model, was effectively developed and proven to achieve accurate polyp detection and segmentation. Based on the models that obtain the best metrics during the training phase, it was possible to find a final model that presented an accuracy of 83.33 %, a precision of 93.90 %, a recall of 82.35 %, an F1 Score of 87.75 %, a dice coefficient of 87.51 % and a medium time of 0.22 seconds per prediction. Finally, the integration evaluation was based on two different approaches: a Flask API and Python.Net. Even though the first one was shown to be a good approach with easy integration, the latency in the data transfer process was fatal for what was proposed. The developed methods and implementations proved that, with some improvements, they could successfully create a polyp detection and segmentation system. |
|---|