Optimizing Blood Clot Detection Using Deep Learning: Effects of Preprocessing Pipelines and Image Slicing Methods

This study explores the use of deep learning models for detecting blood clots in medical images, focusing on how preprocessing and image slicing affect performance. The large size of medical images makes blood clot detection challenging due to memory constraints and the presence of irrelevant background content. We begin by applying preprocessing methods such as image scaling, normalization, and augmentation to make the dataset easier to handle. After that, we train two deep learning models (CNN, DenseNet) and assess how well they perform using a sliced version of the dataset as well as the original high-resolution photos. To address these challenges, we divide images into 1024×1024 patches, which helps reduce memory usage and improve training efficiency. We evaluate model performance using standard metrics such as accuracy and F1-score. Our findings emphasize how data preparation and fragmentation can impact the efficacy of deep learning models in medical image analysis, demonstrating a significant difference in the models' performance when using the entire dataset as opposed to the smaller, divided dataset.