Deep learning detection network for peripheral blood leukocytes based on improved detection transformer

•Pure Transformer architecture outperforms classical CNN models.•PVT can be used as the backbone of DETR to extract multi-scale feature maps.•DAM combined with PVT reduces computational complexity and improves detection performance.•Transfer learning improves the detection performance.•Improved DETR model can be used for leukocyte detection. Leukocytes serve as an important barrier to healthy immunity in the body and play an important role in fighting diseases. Manual morphological examination of leukocytes is the gold standard for the diagnosis of certain diseases but is undoubtedly labour-intensive and requires a high level of expertise. Therefore, conducting research on computer-aided diagnostics is important. With the development of deep learning techniques in computer vision, an increasing number of deep learning-based methods are now being applied in the field of medical imaging. Recently, the detection transformer (DETR) model, which is based on the transformer architecture, has exhibited outstanding performances in object detection tasks and has attracted considerable attention. Our study aims to propose a pure transformer-based end-to-end object detection network based on DETR and apply it to a practical medical scenario of leukocyte detection. First, we introduce the pyramid vision transformer and deformable attention module into the DETR model to improve the model performance and convergence speed. Second, we train the improved model on the challenging Common Objects in Context dataset to obtain the pretrained weights. Third, we perform transfer learning on the modified Raabin leukocyte dataset to obtain the optimal model. The improved DETR shows a mean average precision detection performance of up to 0.961 and is therefore superior to the original DETR and convolutional neural network. The study findings are expected to be useful for the development of a transformer structural model for leukocyte detection.