Abordagem computacional para auxiliar a análise e classificação de câncer de próstata por meio de imagens

Abstract

The analysis of whole-slide histological images (WSI) in prostate cancer diagnosis presents significant computational challenges, such as gigapixel image processing and tissue morphological variability. This work proposes a methodology that integrates a preprocessing stage based on morphological operations with a convolutional neural network architecture of the Mask R-CNN type for instance segmentation. Applied to the PANDA dataset, the approach included the construction of a curated database with patch generation and class balancing. The model, implemented using the Detectron2 framework, achieved an accuracy of 97.87% in classifying Gleason patterns, confirming that targeted preprocessing enhances network learning compared to end-to-end approaches. Validation was also performed using the Karolinska subset. The results indicate the potential of the proposed methodology as a tool for digital pathology, demonstrating its generalization capability and establishing a new reference for applications addressing challenges in medical imaging.