Sistema de visión artificial como apoyo en el proceso de diagnóstico de lesiones cerebrales en la empresa Ecoray Diagnóstico
No hay miniatura disponible
Fecha
2026
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Católica Santo Toribio de Mogrovejo
Resumen
Esta investigación presenta un sistema de visión artificial para apoyar el diagnóstico de lesiones cerebrales en Ecoray Diagnósticos. El objetivo fue integrar un flujo confiable que abarca lectura DICOM, preprocesamiento de tomografías y clasificación automática de subtipos de hemorragia intracraneal, mediante una red DenseNet-121 exportada a ONNX e incorporada en una aplicación de escritorio desarrollada en Java. El modelo se entrenó combinando 3713 estudios públicos rotulados de tomografías cerebrales y 184 estudios locales de la clínica Ecoray. El desarrollo se llevó a cabo con un diseño cuasi experimental y se organizó en módulos. El sistema incluye validación de metadatos DICOM, normalización de las imágenes, reducción de ruido, redimensionamiento y conversión a tensores para la red neuronal, además de un registro de auditoría y la generación de reportes mediante JasperReports. Los resultados muestran una validación basada en 184 estudios locales no utilizados en el entrenamiento para lo cual se obtuvo una exactitud global del 87%, una sensibilidad del 84% y una especificidad del 90%, lo que indica un desempeño clínicamente aceptable en la mayoría de los subtipos modelados los cuales son hemorragia epidural, intraparenquimal, intraventricular, subaracnoidea, subdural y la clase ninguno. El tiempo medio de procesamiento se mantuvo alrededor de cuatro segundos por estudio, reduciendo de forma notable el tiempo necesario para elaborar el informe en comparación con el flujo manual. La usabilidad, medida con el cuestionario SUS aplicado a los radiólogos, alcanzó un 82%, por encima del umbral de 68 considerado aceptable. El sistema se integró al flujo operativo como herramienta de apoyo, sin reemplazar el criterio del especialista, aportando estandarización del preanálisis, priorización de casos y trazabilidad técnica, además de lineamientos para su mantenimiento y mejora futura.
This research presents a computer vision system to support the diagnosis of brain injuries at Ecoray Diagnósticos. The objective was to integrate a reliable workflow that covers DICOM reading, preprocessing of brain computed tomography (CT) scans, and automatic classification of intracranial hemorrhage subtypes, using a DenseNet-121 network exported to ONNX and embedded in a Java desktop application. The model was trained on 3,713 labeled public brain CT studies, while the clinical evaluation focused on local data from the institution. The development followed a quasi-experimental design and was organized in modules. The system includes DICOM metadata validation, image normalization, noise reduction, resizing and conversion to tensors for the neural network, as well as an audit log and automated report generation using JasperReports. The results are based on a validation with 184 local studies that were not used for training, in which the system achieved an overall accuracy of 87%, sensitivity of 84% and specificity of 90%. These values indicate clinically acceptable performance for most of the modeled subtypes: epidural, intraparenchymal, intraventricular, subarachnoid and subdural hemorrhage, plus the “no hemorrhage” class. The average processing time was around four seconds per study, considerably reducing the time required to prepare the report compared with the manual workflow. Usability, measured through the SUS questionnaire applied to radiologists, reached 82%, above the standard acceptability threshold of 68. The system was integrated into the routine workflow as a decision-support tool, without replacing the specialist’s judgment, providing standardized pre-analysis, case prioritization and technical traceability, as well as guidelines for maintenance and future improvement.
This research presents a computer vision system to support the diagnosis of brain injuries at Ecoray Diagnósticos. The objective was to integrate a reliable workflow that covers DICOM reading, preprocessing of brain computed tomography (CT) scans, and automatic classification of intracranial hemorrhage subtypes, using a DenseNet-121 network exported to ONNX and embedded in a Java desktop application. The model was trained on 3,713 labeled public brain CT studies, while the clinical evaluation focused on local data from the institution. The development followed a quasi-experimental design and was organized in modules. The system includes DICOM metadata validation, image normalization, noise reduction, resizing and conversion to tensors for the neural network, as well as an audit log and automated report generation using JasperReports. The results are based on a validation with 184 local studies that were not used for training, in which the system achieved an overall accuracy of 87%, sensitivity of 84% and specificity of 90%. These values indicate clinically acceptable performance for most of the modeled subtypes: epidural, intraparenchymal, intraventricular, subarachnoid and subdural hemorrhage, plus the “no hemorrhage” class. The average processing time was around four seconds per study, considerably reducing the time required to prepare the report compared with the manual workflow. Usability, measured through the SUS questionnaire applied to radiologists, reached 82%, above the standard acceptability threshold of 68. The system was integrated into the routine workflow as a decision-support tool, without replacing the specialist’s judgment, providing standardized pre-analysis, case prioritization and technical traceability, as well as guidelines for maintenance and future improvement.
Descripción
Palabras clave
Diagnóstico, Imágenes médicas, Inteligencia artificial, Diagnosis, Medical imaging, Artificial intelligence
Citación
R. Perez, “Sistema de visión artificial como apoyo en el proceso de diagnóstico de lesiones cerebrales en la empresa Ecoray Diagnóstico,” tesis de licenciatura, Fac. de Ingeniería, Univ. USAT, Chiclayo, Perú, 2026. [En línea]. Disponible en: https://hdl.handle.net/20.500.12423/10115
