A Multi-Model Artificial Intelligence Framework for Automated Blood Clot Detection in Medical Images

Abstract

Pathological clot formation in blood vessels (thrombosis) is a major cause of cardiovascular death throughout the world, accounting for many ischemic strokes, pulmonary embolisms and deep vein thrombosis [9][10]. Timely detection of these lesions in volumetric imaging has a significant impact on a patient's outcome but the massive size of scan acquisitions makes manual radiology assessment an increasingly unrealistic endeavour. We present an AI multi-model framework that integrates a volumetric deep learning segmentation platform with an AI interpreting engine hosted in the cloud to automatically identify, quantify and report clots from CT and MRI scans.

The segmentation component is a 3D U-Net [2][7] model trained with a composite of Dice and Focal losses to address the inherent class imbalances of clot segmentation. A parallel AI vision engine generates narrative clinical reports, including the anatomical location, likely diagnosis, and potential neurological symptoms, which are provided in conjunction with segmentation results for integrated clinical decision making [4][12]. The system is integrated into a Flask web interface that supports drag-and-drop scan upload, real-time inference and visualisation, and report generation in PDF, CSV, and XML. Benchmarking against 2D approaches demonstrate higher Dice Similarity Coefficient and better anatomical coverage of the proposed approach, thus demonstrating its clinical value [1][14].