Modification of YOLO with GhostNet to Reduce Parameters and Computing Resources for Detecting Acute Lymphoblastic Leukemia
Tanzilal Mustaqim, Chastine Fatichah, Nanik Suciati · 2022
The Detection of acute lymphoblastic leukemia (ALL) subtypes on multicellular microscopic images is significant for early diagnosis to support the treatment process. Recently, the object detection with a deep learning approach shows good accuracy and fast computation time in the medical field. Therefore, we propose using the You Only Look Once (YOLO) method, namely the Yolov4 and Yolov5 models, to detect the L1, L2, and L3 subtypes. However, both models still have high GFLOPS values and high number of parameters. This paper proposes a modification of Yolov4 and Yolov5 by replacing the standard backbone convolution module with the GhostNet convolution module. The GhostNet module can reduce the GFLOPS value and the number of parameters. Overall., the Yolo backbone modification model has comparable results with the original Yolo model with a slight difference with a value of 1.4 % in the Yolov4 backbone modification model and 2.4 % in the Yolov5 backbone modification model. The number of parameters and GFLOPS values of the two models modified was reduced by 35% and 40%, respectively.