There is a need to more accurately link human genetic variance with thrombotic risk. Thrombus formation results from adhesion of blood platelets to a site of injury, followed by their progressive aggregation and occasional embolization. To observe this in vitro, blood is perfused over a surface of collagen fibres, during video microscopy. This paper proposes three complementary gradient-based features which, if included in a regularized machine learning framework, yield the accurate segmentation of thrombi during such acquisitions. A novel tracking method of thrombi as deformable growing objects under split and merge conditions is also introduced.