| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.42 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Contexto: A formação médica em cirurgia minimamente invasiva, como a artroscopia, enfrenta
desafios na avaliação de competências técnicas, tradicionalmente baseada na observação
subjetiva de especialistas, através de escalas globais.
Problema: O processo de avaliação é demorado e suscetível a variabilidade entre avaliadores,
comprometendo a consistência e a objetividade das classificações atribuídas.
Métodos: Esta dissertação propõe o desenvolvimento de um sistema baseado em inteligência
artificial para automatizar a avaliação da proficiência cirúrgica em procedimentos de
artroscopia do joelho. A fundamentação teórica foi sustentada por uma revisão sistemática da
literatura, conduzida de acordo com as diretrizes PRISMA. Para o desenvolvimento do sistema,
foi criado um dataset, constituído por vídeos captados através da torre de artroscopia, durante
a execução de exercícios num simulador físico, complementados por anotações de
desempenho, realizadas por especialistas. Estes dados foram utilizados para treinar e validar a
performance do modelo de visão computacional. A abordagem proposta permite avaliar
automaticamente o desempenho com base na escala Arthroscopic Surgical Skill Evaluation Tool
(ASSET), considerando oito dimensões: segurança, campo de visão, destreza da câmara,
destreza com os instrumentos, destreza bimanual, fluxo do procedimento, qualidade do
procedimento e autonomia.
Resultados: As classificações geradas foram comparadas com as avaliações independentes de
dois médicos especialistas. Os resultados demonstraram elevada concordância nas dimensões
de autonomia (88,89 %) e qualidade do procedimento (81,48 %), atingindo 100 % de
concordância dentro de uma margem de um ponto. As dimensões de segurança, destreza com
os instrumentos e fluxo do procedimento apresentaram correspondência exata entre 70 % e
80 %. Verificou-se ainda elevada subjetividade entre avaliadores (α de Krippendorff < 0,67), o
que reforça a necessidade de uma ferramenta de avaliação objetiva.
Conclusão: O sistema de inteligência artificial desenvolvido demonstrou capacidade para
avaliar automaticamente a proficiência cirúrgica em exercícios de artroscopia do joelho,
apresentando níveis elevados de concordância com as avaliações de especialistas, em várias
dimensões da escala ASSET.
Background: Medical training in minimally invasive surgery, such as arthroscopy, faces challenges in the assessment of technical skills, which has traditionally relied on the subjective observation of specialists, using global rating scales. Problem: The assessment process is time-consuming and prone to inter-rater variability, compromising the consistency and objectivity of the ratings assigned. Methods: This thesis proposes the development of an artificial intelligence-based system to automate the assessment of surgical proficiency in knee arthroscopy procedures. The theoretical framework was supported by a systematic review of the literature, conducted in accordance with the PRISMA guidelines. To develop the system, a dataset was created, consisting of videos captured via the arthroscopy tower, whilst performing exercises on a physical simulator, supplemented by performance annotations made by specialists. These data were used to train and validate the performance of the computer vision model. The proposed approach enables automatic assessment of performance based on the Arthroscopic Surgical Skill Evaluation Tool (ASSET) scale, considering eight dimensions: safety, field of view, camera dexterity, instrument dexterity, bimanual dexterity, flow of procedure, quality of procedure and autonomy. Results: The scores generated were compared with the independent assessments of two specialist doctors. The results showed high agreement in the dimensions of autonomy (88,89 %) and quality of the procedure (81,48 %), with 100 % agreement within a margin of one point. The dimensions of safety, instrument dexterity and procedural flow showed exact agreement between 70 % and 80 %. High subjectivity was also observed among assessors (Krippendorff’s α < 0,67), which reinforces the need for an objective assessment tool. Conclusion: The artificial intelligence system developed demonstrated the ability to automatically assess surgical proficiency in knee arthroscopy exercises, showing high levels of agreement with expert assessments across various dimensions of the ASSET scale.
Background: Medical training in minimally invasive surgery, such as arthroscopy, faces challenges in the assessment of technical skills, which has traditionally relied on the subjective observation of specialists, using global rating scales. Problem: The assessment process is time-consuming and prone to inter-rater variability, compromising the consistency and objectivity of the ratings assigned. Methods: This thesis proposes the development of an artificial intelligence-based system to automate the assessment of surgical proficiency in knee arthroscopy procedures. The theoretical framework was supported by a systematic review of the literature, conducted in accordance with the PRISMA guidelines. To develop the system, a dataset was created, consisting of videos captured via the arthroscopy tower, whilst performing exercises on a physical simulator, supplemented by performance annotations made by specialists. These data were used to train and validate the performance of the computer vision model. The proposed approach enables automatic assessment of performance based on the Arthroscopic Surgical Skill Evaluation Tool (ASSET) scale, considering eight dimensions: safety, field of view, camera dexterity, instrument dexterity, bimanual dexterity, flow of procedure, quality of procedure and autonomy. Results: The scores generated were compared with the independent assessments of two specialist doctors. The results showed high agreement in the dimensions of autonomy (88,89 %) and quality of the procedure (81,48 %), with 100 % agreement within a margin of one point. The dimensions of safety, instrument dexterity and procedural flow showed exact agreement between 70 % and 80 %. High subjectivity was also observed among assessors (Krippendorff’s α < 0,67), which reinforces the need for an objective assessment tool. Conclusion: The artificial intelligence system developed demonstrated the ability to automatically assess surgical proficiency in knee arthroscopy exercises, showing high levels of agreement with expert assessments across various dimensions of the ASSET scale.
Descrição
Palavras-chave
Artroscopia Simulação Cirúrgica Inteligência Artificial Visão Computacional Análise de Vídeos Proficiência Cirúrgica Arthroscopy Surgical simulation Artificial Intelligence Computer vision Video analysis Surgical proficiency