Medical image classification based on representational learning

• Autores: Juan Antonio Castro Silva
• Directores de la Tesis: María Navelonga Moreno García (dir. tes.)
• Lectura: En la Universidad de Salamanca ( España ) en 2025
• Idioma: inglés
• Tribunal Calificador de la Tesis: Manuel Martín-Merino Acera (presid.) , Vivian Félix López Batista (secret.) , Carlos Alberto Cobos Lozada (voc.)
• Enlaces
  • Tesis en acceso abierto en: TESEO
• Resumen

  • Accurate and early diagnosis of Alzheimers Disease (AD) remains a significant challenge due to the complexity of its pathological progression and the subtlety of its early biomarkers. Structural magnetic resonance imaging (MRI), combined with advancements in deep learning, has shown promising results for AD classification. However, existing approaches suffer from limitations such as fixed or heuristic slice selection, poor region of interest (ROI) targeting, and inadequate handling of anatomical variability. These shortcomings can result in data redundancy, reduced model generalizability, and increased computational costs.

    This thesis addresses these limitations by proposing a novel, ROI-guided slice instance selection methodology that integrates multi-atlas information to improve the representativeness and informativeness of input data for AD classification. A statistical centroid-based ROI extraction method is also introduced to localize and crop disease-relevant image regions precisely. The selected 2D slices and ROI patches are further evaluated using deep convolutional neural networks (CNNs) and hybrid ensemble methods to assess classification performance across anatomical planes, preprocessing variations, and CNN architectures. Additionally, a multiple-input, mixed-data 3D Vision Transformer (ViT) ensemble model is presented to incorporate multimodal data, combining 3D MRI with demographic and cognitive scores, to improve diagnostic accuracy.

    The proposed methods were validated using three large-scale public datasets (ADNI, AIBL, and OASIS), and the results demonstrate statistically significant improvements over both the baseline and state-of-the-art models. The hybrid ensemble achieved a maximum classification accuracy of 95%, and the proposed 3D ViT outperformed comparable architectures in multiple configurations. These contributions highlight the effectiveness of anatomically informed instance selection and the value of hybrid and multimodal deep learning approaches for robust and scalable AD diagnosis.