LAFSSD: lightweight and advanced FSSD for multi-scale detection of platelets and white blood cells in human peripheral blood smear images.

Accurate real-time detection of platelets and white blood cells (WBCs) is crucial to the diagnosis of clinical disease and blood cell classification. This paper proposes a lightweight and advanced feature fusion single shot multibox detector (LAFSSD) model to enhance the multi-scale detection capability of platelets and WBCs. To generate a more effective multi-scale feature fusion module composed of depthwise separable convolution, the model first substitutes VGG16, the basic backbone of the FSSD model, with a lightweight MobileNet network. Afterward, the K-means clustering algorithm is used to modify the aspect ratio and size of the default box. The CBAM attention modules in the multi-scale feature map of LAFSSD's structure are added to further enhance the detection of platelets and WBCs. Finally, the RMSProp gradient descent algorithm is used to decrease the irregular oscillations of the loss function. The experimental outcomes demonstrate that the mAP of the LAFSSD model is 98.4%, a rise of 24.7% from before the enhancement. Our model can precisely and in real-time detect platelets and WBCs at multiple scales of different sizes and shapes in human peripheral blood smear images. [ABSTRACT FROM AUTHOR]