Thesis Plataforma web para la detección y clasificación automatizada de lesiones en mamografías mediante aprendizaje profundo
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Telemática
Departament
Campus
Campus Santiago San Joaquín
Abstract
Este trabajo presenta el desarrollo de una plataforma web integral para la detección y clasificación automatizada de lesiones en mamografías, combinando métodos avanzados de procesamiento de imágenes médicas, aprendizaje profundo y visualización interactiva. A partir de un análisis exploratorio de datos, se realizaron etapas de preprocesamiento minucioso de imágenes DICOM, extrayendo regiones de interés (ROI) y aplicando técnicas de aumento de datos. Para la clasificación se empleó Vision Transformer (ViT), logrando discriminar entre masas, calcificaciones y otras lesiones con buenos resultados. Paralelamente, se integró un modelo de detección basado en YOLOX para resaltar visualmente las lesiones más sospechosas.
La solución se consolido en una plataforma web desarrollada con Streamlit, facilitando la carga, procesamiento masivo de imágenes y generación de reportes personalizados en PDF. Esta herramienta resulta intuitiva para el usuario final, reduciendo la carga de trabajo del especialista y permitiendo priorizar los casos con mayor riesgo, contribuyendo así a la detección temprana del cancer de mama. Finalmente, se discuten limitaciones del modelo, implicancias clínicas, posibles mejoras técnicas y futuras validaciones clínicas, con el objetivo de fomentar la adopción efectiva de la inteligencia artificial en entornos radiológicos reales.
This work presents the development of a comprehensive web platform for the automated detection and classification of lesions in mammograms, combining advanced methods of medical image processing, deep learning and interactive visualization. Starting from an exploratory data analysis, thorough DICOM image preprocessing steps were performed, extracting regions of interest (ROI) and applying data augmentation techniques. Vision Transformer (ViT) was used for classification, discriminating between masses, calcifications and other lesions with good results. In parallel, a YOLOX-based detection model was integrated to visually highlight the most suspicious lesions. The solution was consolidated in a web platform developed with Streamlit, facilitating the loading, massive image processing and generation of customized PDF reports. This tool is intuitive for the end user, reducing the workload of the specialist and allowing to prioritize the cases with higher risk, thus contributing to the early detection of breast cancer. Finally, limitations of the model, clinical implications, possible technical improvements and future clinical validations are discussed, with the aim of promoting the effective adoption of artificial intelligence in real radiological environments.
This work presents the development of a comprehensive web platform for the automated detection and classification of lesions in mammograms, combining advanced methods of medical image processing, deep learning and interactive visualization. Starting from an exploratory data analysis, thorough DICOM image preprocessing steps were performed, extracting regions of interest (ROI) and applying data augmentation techniques. Vision Transformer (ViT) was used for classification, discriminating between masses, calcifications and other lesions with good results. In parallel, a YOLOX-based detection model was integrated to visually highlight the most suspicious lesions. The solution was consolidated in a web platform developed with Streamlit, facilitating the loading, massive image processing and generation of customized PDF reports. This tool is intuitive for the end user, reducing the workload of the specialist and allowing to prioritize the cases with higher risk, thus contributing to the early detection of breast cancer. Finally, limitations of the model, clinical implications, possible technical improvements and future clinical validations are discussed, with the aim of promoting the effective adoption of artificial intelligence in real radiological environments.
Description
Keywords
Plataforma web, Mamografía, Detección automatizada, Diagnóstico asistido por computadores