Browsing by Author "Michelucci, Umberto"
Now showing 1 - 2 of 2
Results Per Page
Sort Options
Item Histology Tissue Classification for Colorectal Cancer: Deep Learning Approaches for Medical Image Analysis. Which deep learning architecture provides the best balance between accuracy and computational efficiency for histological tissue classification in colorectal cancer?(Universidad del Desarrollo. Facultad de Ingeniería, 2025) Castro Ortega, Kurt; Gómez Vargas, Germán; Michelucci, UmbertoThis report presents a comprehensive comparative analysis of deep learning architectures for the classification of histological images of colorectal cancer, a fundamental task in the emerging field of digital pathology. Automating this process through the use of deep learning models not only promises to accelerate diagnosis, but also to improve its accuracy and consistency, representing a strategic opportunity to optimize clinical workflows and advance toward precision medicine. The project addressed the classification of eight types of colorectal cancer tissue through a systematic comparative evaluation of five deep learning architectures: a baseline Convolutional Neural Network (CNN) trained from scratch and four advanced models (VGG19, ResNet50, EfficientNetB0, and Vision Transformer ViT-B16) implemented with the Transfer Learning strategy. The results showed that the ResNet50 architecture achieved the highest performance, with an accuracy of 95.07% on the test set. The analysis validated the project's central hypothesis: pre-trained architectures significantly outperformed the model trained from scratch, underscoring the effectiveness of transfer learning in domains with limited data. Furthermore, the study revealed a complex balance between performance and computational efficiency, where theoretically more efficient models such as EfficientNetB0 do not always translate into the lowest inference latency on specific hardware. This report provides a detailed analysis of these findings, offering empirical guidance for architecture selection in computational pathology applications.Item XAI in CNNS for medical use: classification of colorectal cancer histological textures through fine-tuning of kimianet and comparative analysis coof explainability(Universidad del Desarrollo. Facultad de Ingeniería, 2025) Lavarello Osorio, Giuseppe; Gómez Vargas, Germán; Michelucci, UmbertoColorectal cancer (CRC) is characterized by significant tissue heterogeneity, making histopathological analysis a critical but complex task for diagnosis and prognosis. This project addresses the need for an automated classification system that is not only accurate but also interpretable, the most crucial factor for building trust and facilitating adoption in clinical settings. The proposed solution is twofold. First, it involves the fine-tuning of a domain-specific DenseNet-121 model, pre-trained with the vast KimiaNet histopathology image repository, on the public Kather et al. (2016) dataset of CRC textures. Second, to address the "black box" nature of Deep Learning models that limits their clinical utility, this study conducts a rigorous comparative analysis of advanced Explainable AI (XAI) techniques, including Grad-CAM, XGrad-CAM, and SmoothGrad. This analysis validates the model's reliability through both qualitative visual inspection and quantitative fidelity metrics. The central argument of this project is that a clinically useful classifier must be both accurate and trustworthy; this work presents a complete workflow to build and verify that trust.