Programa del congreso
| Sesión | ||||||
Mi-S1.1-IAIM1: Inteligencia artificial en imagen médica (I)
| ||||||
| Ponencias | ||||||
11:45 - 12:00
Priorización de lesiones cutáneas: cómo los sistemas de IA fallan cuando se prueban utilizando bases de datos de escenarios reales 1Universidad de Sevilla, España; 2Hospital Universitario Virgen Macarena, Sevilla, España En los últimos años, el número de sistemas de inteligencia artificial (IA) desarrollados para la clasificación de lesiones cutáneas ha crecido de manera exponencial, impulsado en gran parte por la disponibilidad de bases de datos públicas. Sin embargo, muchos de estos modelos presentan limitaciones al desplegarse en entornos clínicos reales, donde variaciones en la calidad de los datos afectan a su capacidad de generalización. En esta investigación se propone un sistema basado en la reinterpretación de la tarea tradicional de clasificación hacia la priorización clínica: en lugar de predecir diagnósticos específicos, categorizar las lesiones en tres niveles en relación con la urgencia a la hora de recibir diagnóstico. El modelo desarrollado se basa en una ConvNeXt Tiny con un clasificador final y se implementan técnicas como focal loss y ponderación de clases para abordar el desbalance de clases. Asimismo, diversos sistemas de IA reportados en la literatura son analizados en profundidad para finalmente seleccionar y adaptar dos a la tarea de priorización en tres niveles. Aunque en sus contextos originales mostraron buen rendimiento, en el escenario clínico con datos reales propuesto, presentan una notable caída en la precisión. En contraste, el modelo propuesto alcanza mejoras significativas de entre el 13% y el 21% para el mismo escenario evaluado. Estos resultados subrayan la importancia de adaptar las herramientas de IA a tareas de triaje en contextos clínicos reales y evidencia el potencial del sistema propuesto como apoyo en la toma de decisiones por parte de los profesionales.
12:00 - 12:15
Deep Learning-based age prediction models from retinal Optical Coherence Tomography images 1Mondragon Goi Eskola Politeknikoa, España; 2Instituto de Investigación Sanitaria Biobizkaia; 3Ikerbasque, Fundación Vasca para la Ciencia This study evaluates the potential of Optical Coherence Tomography (OCT) as a non-invasive tool for retinal age prediction in healthy individuals. A dataset comprising 1,180 eyes from 517 control subjects was used to compare deep learning models trained on different OCT scan types: peripapillary B-scans, individual macula raster B-Scans, and full macular volumes. Images underwent standardized preprocessing, and models based on 2D and 3D ResNet architectures were trained and optimized using Transfer Learning. Results show that volumetric macular scans applied in a ResNet-3D model achieved the lowest Mean Absolute Error (3.07 years), outperforming both previous literature and all tested 2D configurations. Overall, findings highlight that integrating depth and spatial features in OCT data significantly enhances retinal age estimation.
12:15 - 12:30
Zero-shot segmentation of skin tumors in whole-slide images with vision-language foundation models 1Instituto Universitario de Investigación en Tecnología Centrada en el Ser Humano (HUMAN-Tech), Universitat Politecnica de Valencia (UPV), Valencia, España; 2Artikode Intelligence S.L., Valencia, España Accurate annotation of cutaneous neoplasm biopsies represents a major challenge due to their wide morphological variability, overlapping histological patterns, and the subtle distinctions between benign and malignant lesions. Vision–language foundation models (VLMs), pre‐trained on paired image–text corpora, learn joint representations that bridge visual features and diagnostic terminology, enabling zero‐shot localization and classification of tissue regions without pixel‐level labels. However, most existing VLM applications in histopathology remain limited to slide‐level tasks or rely on coarse interactive prompts, and they struggle to produce fine‐grained segmentations across gigapixel whole‐slide images (WSIs). In this work, we introduce a Zero-shot visual-language segmentation pipeline for whole-slide images (ZEUS), a fully automated, zero‐shot segmentation framework that leverages class‐specific textual prompt ensembles and frozen VLM encoders to generate high‐resolution tumor masks in WSIs. By partitioning each WSI into overlapping patches, extracting visual embeddings, and computing cosine similarities against text prompts in order to generate a final segmentation mask. We demonstrate competitive performance on two in‐house datasets, primary spindle cell neoplasms and cutaneous metastases, highlighting the influence of prompt design, domain shifts, and institutional variability in VLMs for histopathology. ZEUS markedly reduces annotation burden while offering scalable, explainable tumor delineation for downstream diagnostic workflows.
12:30 - 12:45
Clasificación automática de psoriasis empleando EfficientNetB0 y transferencia de aprendizaje Universidad de Sevilla, España La psoriasis es una enfermedad inflamatoria crónica con múltiples variantes clínicas cuya identificación temprana resulta clave para un manejo clínico adecuado. En este trabajo presentamos un sistema automático de clasificación de imágenes dermatológicas de psoriasis capaz de distinguir entre cinco de sus variantes (placas, guttata, inversa, pustulosa y eritrodérmica). Para ello, se creó una base de datos híbrida de 995 imágenes, integrando dos fuentes públicas (DermNet y Roboflow Universe) y 54 casos clínicos procedentes de una base de datos privada con imágenes hospitalarias. Sobre este conjunto de datos se aplicó aprendizaje por transferencia usando EfficientNetB0 como extractor de características y un clasificador personalizado con capas densas, batch normalization y dropout. Para reducir el efecto del desbalance entre clases se implementó una estrategia de data augmentation y balanceo. La validación se realizó con un protocolo K-fold estratificado con k = 5 y evaluaciones Leave-One-Out (LOOCV) complementarias. El modelo final alcanzó una precisión media del 92.52%, con un comportamiento estable entre los distintos folds. En este trabajo se presenta una comparativa con otro trabajo reciente en la misma línea de investigación y con resultados del estado del arte en clasificación. Además, se ha realizado una comparación con otras estrategias de aprendizaje profundo en términos de exactitud y tiempos de computación, mostrando las ventajas de la estrategia elegida para poder ser empleada y reproducida en equipos sin necesidades especiales de equipos hardware complejos para computación.
12:45 - 13:00
Deep Learning for Breast Cancer Screening Using Diffusion-Weighted MRI 1Computer Science Department, University of Oviedo, Spain; 2Electrical Engineering Department, University of Oviedo, Spain; 3Biomedical Engineering Center, University of Oviedo, Spain; 4School of Mechanical Engineering, Purdue University, 585 Purdue Mall, West Lafayette, IN 47907, USA; 5Purdue Center for Cancer Research, Purdue University, West Lafayette, IN, USA; 6Weldon School of Biomedical Engineering, Purdue University, 206 S. Martin Jischke Drive, West Lafayette, IN 47907, USA; 7Group of Numerical Methods in Engineering, Department of Mathematics, Campus de Elvi˜na s/n, 15071 A Coru˜na, Galicia, Spain; 8Oden Institute for Computational Engineering and Sciences, The University of Texas, 201 E. 24th Street, Austin, TX 78712-1229, USA; 9Radiodiagnostic Service, Oviedo Central University Hospital (HUCA), Oviedo, Spain Cancer remains a leading cause of global mortality, with breast cancer showing particularly high incidence among women. Early detection significantly improves prognosis. Diagnosis and monitoring are commonly performed either using mammography or dynamic contrast-enhanced magnetic resonance imaging. Recently, diffusion-weighted magnetic resonance imaging has gathered increased attention as an alternative to both techniques due to some key advantages, although it has not yet achieved widespread use in clinical workflows, as diffusion weighted images (DWI) present limitations that harden the extraction of information from them. Recent advances in artificial intelligence, especially in the field of computer vision, offer opportunities to improve and automate information extraction from DWI. In this study, we explore the use of DWI in screening workflows with the creation and evaluation of a set of classification algorithms. The results obtained by our algorithms display the potential of deep learning applied to DWI, achieving an F1 score of 84% and a recall of nearly 90%.
13:00 - 13:15
Evaluation of Hip Dysplasia using Deep Learning for the Automated Detection of Wiberg and Tönnis angles in X-rays 1Master in Biomechanical Engineering and Medical Devices, Universidad Carlos III de Madrid, Spain; 2Hip Unit, Clínica CEMTRO, Madrid, Spain Hip dysplasia is an orthopedic condition associated with an atypically shaped acetabulum, leading to hip instability. The primary tool for assessing acetabular coverage is the x-ray once ossification of the femoral head has commenced. Key indicators such as Tönnis and Wiberg angles can be measured, which will be the focus of this study. These angles are currently measured manually on radiographs, which introduces variability, motivating the development of automated methods. Therefore, we propose BAGGY©, a deep learning-based software for automatic detection of anatomical landmarks and measurement of Wiberg and Tönnis angles. The system employs a Faster Region-based Convolutional Neural Network (Faster R-CNN) with a Residual Network-50 (ResNet-50) and a Feature Pyramid Network (FPN) backbone, pretrained on Common Objects in Context (COCO), fine-tuned with custom detection heads for landmark detection. We collected 101 radiographs for training, validation and testing. A 5-fold cross-validation method was employed in which landmark detection achieved a mean average precision of 0.76 for both angles. On the independent test set annotated by a doctor, deviations in landmark predictions averaged ~ 5 pixels. Angle measurements showed strong agreement with the doctor (Pearson’s correlation coefficient r~0.85, intraclass correlation coefficient ICC>0.8), with the Tönnis angle demonstrating higher precision (2.9º) than the Wiberg angle (5.0º). Automatic angle computation was feasible in most radiographs (76.9% for Wiberg, and 84.6% for Tönnis). These results support the potential applicability of the proposed approach to assist in the evaluation of hip dysplasia
| ||||||