Adaptive elastic convolution-based YOLO for peripheral blood smear cell detection

Abstract

Peripheral blood smear analysis continues to be important for the diagnosis of various hematologic conditions, although automated systems often experience difficulties due to the variety of cell types and the multi-class detection requirements associated with them. In this work, we propose Elastic YOLO (EYOLO), an extension of the You Only Look Once (YOLO) object detection framework designed for morphology-aware detection of red blood cells, white blood cells, and platelets in peripheral blood smear images. In a dataset annotated by clinicians, Elastic YOLO achieved a mAP@0.5 of 94.7% and mAP@0.5:0.95 of 87.8%, outperforming the baseline YOLOv5 model and several recent detection architectures while achieving inference speeds of up to 78 frames per second (FPS) under high-performance GPU settings (NVIDIA RTX 4090) at 256 × 256 input resolution. Elastic adaptive convolutions form the core of the framework, allowing the receptive field to adapt dynamically to variations in cell size, shape, and staining conditions across different cell types. The automated smear image analysis offered by this system may assist computer-assisted hematology workflows and can support remote screening scenarios such as telemedicine-based triage, where rapid preliminary analysis of peripheral smear images is required.