A SAM-Based Automated Schistocyte Detection Pipeline in Peripheral Blood Smear Images

The diagnosis of haematological diseases often relies on accurately identifying specific cell types in the peripheral blood smear (PBS), especially when diagnosing rare diseases. This study focuses on supporting the diagnosis of Thrombotic Thrombocytopenic Purpura (TTP), a life-threatening condition for which detecting schistocytes in PBS images is essential. Schistocytes, irregularly shaped and fragmented red blood cells (RBCs), represent a critical diagnostic marker, making their accurate localisation a fundamental step in the diagnostic process. The complexity of this task lies in the scarcity of available data and the subtle morphological difference between schistocytes and healthy RBCs. Given the limited studies currently available in the literature, in this work, we investigated a set of machine learning approaches to establish an initial baseline for the task of schistocyte identification. The techniques explored include object detection, instance segmentation, as well as a pipeline combining a general-purpose segmentation model with additional steps designed explicitly for schistocyte recognition. Although the current results are not yet sufficient to support the reliable use of the tool for TTP diagnosis, this study offers a solid and comprehensive foundation for future developments, providing valuable insights and benchmarks for this underexplored task.