Thesis Propuesta de diseño de un modelo de machine learning para la clasificación de melanomas en la piel, a partir de imágenes dermatoscopicas
No Thumbnail Available
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Informática
Departament
Campus
Campus Santiago San Joaquín
Abstract
El estudio de lesiones en la piel mediante dermatoscopia es crucial para la detección temprana de melanomas. Incorporar apoyo al proceso con herramientas tecnológicas que no provee la informática, es un gran paso en mejorar de los procedimientos. En este trabajo, se propone un modelo basado en redes neuronales capaz de clasificar ocho tipos de lesiones. El modelo utiliza un enfoque jerárquico para abordar el desbalance de clases. Si bien la propuesta presenta un valor de recall similar comparado con el estado del arte [Abdelhafeez et al., 2023] (75.03 vs 76.1), los resultados en accuracy y precisión son inferiores (74.19 y 61.69 vs. 85.74 y 84) respectivamente. A partir del estudio de los resultados, se proponen mejoras para optimizar el rendimiento del modelo.
The study of skin lesions through dermatoscopy is crucial for the early detection of melanomas. Incorporating support to the process with technological tools provided by computing is a big step in improving procedures. In this work, a model based on neural networks capable of classifying eight types of injuries is proposed. The model use a hierarchical approach to attack class imbalance. Although the proposal presents a similar recall value compared to the state-of-the-art [Abdelhafeez et al., 2023] (75.03 vs. 76.1), the results in terms of accuracy and precision are lower (74.19 and 61.69 vs. 85.74 and 84), respectively. Based on the results study, i propose improvements to optimize the model’s performance.
The study of skin lesions through dermatoscopy is crucial for the early detection of melanomas. Incorporating support to the process with technological tools provided by computing is a big step in improving procedures. In this work, a model based on neural networks capable of classifying eight types of injuries is proposed. The model use a hierarchical approach to attack class imbalance. Although the proposal presents a similar recall value compared to the state-of-the-art [Abdelhafeez et al., 2023] (75.03 vs. 76.1), the results in terms of accuracy and precision are lower (74.19 and 61.69 vs. 85.74 and 84), respectively. Based on the results study, i propose improvements to optimize the model’s performance.
Description
Keywords
Redes neuronales, Clasificación jerárquica, Melanoma, Dermatoscopia, Desbalance de clases
