Thesis Propuesta de modelo de ML que apoye la detección temprana de cáncer bucal a través de técnicas de aprendizaje automático
Loading...
Date
2025-03
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Informática
Departament
Campus
Campus Casa Central Valparaíso
Abstract
El cáncer se ha consolidado como la segunda causa de muerte en Chile, y el diagnóstico tardío es un factor determinante que reduce significativamente la sobrevida del paciente. En particular, se prevé que la incidencia del cáncer oral aumente en los próximos años. Esta enfermedad, de carácter silencioso, suele detectarse en etapas avanzadas, cuando ya ha hecho metástasis, lo que disminuye drásticamente las posibilidades de tratamiento efectivo y recuperación.
Hasta la fecha, los métodos tradicionales de diagnóstico —como los clínicos, ópticos, histológicos y las biopsias— han sido las principales herramientas utilizadas para la detección precisa de esta patología. Sin embargo, con el avance de la tecnología y el desarrollo de técnicas de Machine Learning, han surgido nuevas alternativas que permiten optimizar y agilizar los procesos diagnósticos mediante el análisis de imágenes médicas. Los estudios revisados en el estado del arte han explorado distintas fuentes de imágenes, incluyendo muestras de tejido obtenidas a través de biopsias y registros visuales de zonas específicas de la cavidad oral, como la lengua.
En este trabajo, se adoptó un enfoque más generalizable al utilizar imágenes de diversas secciones de la cavidad bucal, abarcando no solo la lengua, sino también mejillas, paladar y labios, en conjunto con la implementación de un ensamble de modelos (ResNet50, Xception e InceptionV3). Esto permite desarrollar modelos más robustos capaces de detectar patrones asociados a lesiones malignas en distintas áreas de la boca. Los resultados obtenidos destacan el potencial del enfoque propuesto, alcanzando un F1-score de 0,878 y un AUC-ROC de 0,87.
Estos resultados demuestran la viabilidad del aprendizaje automático como una herramienta complementaria en el diagnóstico médico, con el potencial de integrarse en los servicios de salud y contribuir significativamente a la detección temprana del cáncer oral.
Cancer has become the second leading cause of death in Chile, with late diagnosis being a critical factor that significantly reduces patient survival. In particular, the incidence of oral cancer is expected to increase in the coming years. This silent disease is often detected at advanced stages when metastasis has already occurred, drastically reducing the chances of effective treatment and recovery. To date, traditional diagnostic methods—such as clinical, optical, histological techniques, and biopsies—have been the primary tools used for the accurate detection of this pathology. However, with technological advancements and the development of Machine Learning techniques, new alternatives have emerged to optimize and accelerate diagnostic processes through medical image analysis. The studies reviewed in the state of the art have explored various image sources, including tissue samples obtained through biopsies and visual records of specific areas of the oral cavity, such as the tongue. This study adopts a more generalizable approach by utilizing images from various sections of the oral cavity, covering not only the tongue but also the cheeks, palate, and lips. Additionally, it implements an ensemble of models (ResNet50, Xception, and InceptionV3). This approach enables the development of more robust models capable of detecting patterns associated with malignant lesions in different areas of the mouth. The results highlight the potential of the proposed approach, achieving an F1-score of 0,878 and an AUC-ROC of 0,87. These findings demonstrate the feasibility of Machine Learning as a complementary tool in medical diagnostics, with the potential to integrate into healthcare services and make a significant contribution to the early detection of oral cancer.
Cancer has become the second leading cause of death in Chile, with late diagnosis being a critical factor that significantly reduces patient survival. In particular, the incidence of oral cancer is expected to increase in the coming years. This silent disease is often detected at advanced stages when metastasis has already occurred, drastically reducing the chances of effective treatment and recovery. To date, traditional diagnostic methods—such as clinical, optical, histological techniques, and biopsies—have been the primary tools used for the accurate detection of this pathology. However, with technological advancements and the development of Machine Learning techniques, new alternatives have emerged to optimize and accelerate diagnostic processes through medical image analysis. The studies reviewed in the state of the art have explored various image sources, including tissue samples obtained through biopsies and visual records of specific areas of the oral cavity, such as the tongue. This study adopts a more generalizable approach by utilizing images from various sections of the oral cavity, covering not only the tongue but also the cheeks, palate, and lips. Additionally, it implements an ensemble of models (ResNet50, Xception, and InceptionV3). This approach enables the development of more robust models capable of detecting patterns associated with malignant lesions in different areas of the mouth. The results highlight the potential of the proposed approach, achieving an F1-score of 0,878 and an AUC-ROC of 0,87. These findings demonstrate the feasibility of Machine Learning as a complementary tool in medical diagnostics, with the potential to integrate into healthcare services and make a significant contribution to the early detection of oral cancer.
Description
Keywords
Redes neuronales, Prevención de enfermedades, Automatización
