Publication: Relación entre las representaciones generadas por un modelo de inteligencia artificial entrenado con imágenes dermatoscópicas y las características clínicas vinculadas al índice de Breslow en melanomas
Authors
Alarcón Soldevilla, Fernando
item.page.secondaryauthor
Escuela Internacional de Doctorado
item.page.director
Pereyra Rodríguez, José Juan ; Alonso Romero, José Luis
Publisher
Universidad de Murcia
publication.page.editor
publication.page.department
DOI
item.page.type
info:eu-repo/semantics/doctoralThesis
Description
Abstract
La dermatoscopia es una herramienta esencial para el diagnóstico del melanoma, ya que permite visualizar estructuras no apreciables a simple vista. A pesar de ello, existe una considerable variabilidad interobservador, especialmente en lesiones con patrones ambiguos. En este contexto, la inteligencia artificial (IA) -en particular las redes neuronales convolucionales (RNC)- ha surgido como una alternativa útil para analizar imágenes dermatoscópicas mediante algoritmos de clasificación. No obstante, la mayoría de los trabajos se han centrado en diagnósticos dicotómicos, con menor atención a su relación con parámetros pronósticos como el índice de Breslow (IB).El objetivo de este estudio fue evaluar la relación entre las representaciones generadas por un modelo de IA entrenado con imágenes dermatoscópicas de melanomas y las características clínicas asociadas al grosor tumoral. Se diseñó un estudio observacional y correlacional utilizando 1162 imágenes, de las cuales 720 contaban con medición histológica precisa del IB. El modelo empleado fue ConvNext-small, entrenado para diferenciar melanomas superficiales (<0,76 mm) y profundos (###0,76 mm).El análisis se estructuró en tres fases: (1) correlación entre las probabilidades del modelo y el grosor tumoral mediante regresión lineal; (2) reducción dimensional de las representaciones internas mediante Análisis de Componentes Principales (ACP); y (3) validación clínica de los agrupamientos latentes del ACP en relación con patrones dermatoscópicos y categorías de Breslow, empleando chi-cuadrado, test de Fisher-Freeman-Halton y V de Cramer.El modelo mostró una correlación positiva entre sus probabilidades y el IB (R2 = 0,25). El ACP reveló una organización progresiva del espacio latente en función del grosor, con mayor variabilidad en el intervalo intermedio. Asimismo, los agrupamientos se alinearon con patrones dermatoscópicos reconocidos: policromía y estructuras blanco-azuladas en melanomas invasivos, y áreas de regresión, retículo atípico y glóbulos atípicos en melanomas superficiales.Estos hallazgos sugieren que la IA puede capturar patrones estructurales relacionados con el grosor tumoral, constituyendo una herramienta complementaria en situaciones de incertidumbre diagnóstica.
Dermoscopy is an essential tool for melanoma diagnosis, allowing visualization of structures not visible to the naked eye. However, considerable interobserver variability persists, particularly in lesions with ambiguous or atypical patterns. In this setting, artificial intelligence (AI), especially convolutional neural networks (CNNs), has emerged as a useful alternative for analyzing dermoscopic images through classification algorithms. Nevertheless, most studies have focused on binary diagnostic tasks, with limited exploration of their relationship to prognostic indicators such as Breslow thickness (BT). The aim of this study was to assess the relationship between the representations generated by an AI model trained on dermoscopic melanoma images and the clinical features associated with tumor thickness. An observational, correlational study was conducted using 1,162 images, of which 720 had an accurate histological BT measurement. The model used was ConvNext-small, trained to distinguish between superficial (<0.76 mm) and deep (>=0.76 mm) melanomas. The analysis was structured into three phases: (1) evaluation of the correlation between the model's predicted probabilities and tumor thickness using linear regression; (2) dimensionality reduction of internal representations via Principal Component Analysis (PCA); and (3) clinical validation of the PCA-derived latent clusters in relation to dermoscopic patterns and Breslow categories, using chi-square tests, the Fisher-Freeman-Halton test, and Cramer's V. The model showed a positive correlation between its predicted probabilities and BT (R2 = 0.25). PCA revealed a progressive organization of the latent space according to tumor thickness, with greater variability in the intermediate group. The clusters also aligned with clinically recognized dermoscopic patterns: polychromia and blue-white structures in invasive melanomas, and regression areas, atypical networks, and atypical globules in superficial melanomas. These findings suggest that AI can capture structural patterns related to tumor thickness, offering a complementary tool in scenarios of diagnostic uncertainty.
Dermoscopy is an essential tool for melanoma diagnosis, allowing visualization of structures not visible to the naked eye. However, considerable interobserver variability persists, particularly in lesions with ambiguous or atypical patterns. In this setting, artificial intelligence (AI), especially convolutional neural networks (CNNs), has emerged as a useful alternative for analyzing dermoscopic images through classification algorithms. Nevertheless, most studies have focused on binary diagnostic tasks, with limited exploration of their relationship to prognostic indicators such as Breslow thickness (BT). The aim of this study was to assess the relationship between the representations generated by an AI model trained on dermoscopic melanoma images and the clinical features associated with tumor thickness. An observational, correlational study was conducted using 1,162 images, of which 720 had an accurate histological BT measurement. The model used was ConvNext-small, trained to distinguish between superficial (<0.76 mm) and deep (>=0.76 mm) melanomas. The analysis was structured into three phases: (1) evaluation of the correlation between the model's predicted probabilities and tumor thickness using linear regression; (2) dimensionality reduction of internal representations via Principal Component Analysis (PCA); and (3) clinical validation of the PCA-derived latent clusters in relation to dermoscopic patterns and Breslow categories, using chi-square tests, the Fisher-Freeman-Halton test, and Cramer's V. The model showed a positive correlation between its predicted probabilities and BT (R2 = 0.25). PCA revealed a progressive organization of the latent space according to tumor thickness, with greater variability in the intermediate group. The clusters also aligned with clinically recognized dermoscopic patterns: polychromia and blue-white structures in invasive melanomas, and regression areas, atypical networks, and atypical globules in superficial melanomas. These findings suggest that AI can capture structural patterns related to tumor thickness, offering a complementary tool in scenarios of diagnostic uncertainty.
publication.page.subject
Citation
item.page.embargo
Collections
Ir a Estadísticas
Este ítem está sujeto a una licencia Creative Commons. http://creativecommons.org/licenses/by-nc-nd/4.0/