A lightweight YOLOv8-based model for gastric cancer detection

Recent research on deep learning-based gastric cancer detection has demonstrated high performance, with capabilities comparable to or exceeding those of medical professionals. However, the performance of deep learning models depends on the performance of processors such as GPU and CPU, and real-world medical environments encompass a diverse array of computer processor. Recognizing the necessity of research that accommodates varying processors, this study proposes a gastric cancer detection model based on YOLOv8, aiming to achieve real-time performance with reduced sensitivity to performance fluctuations of computer environments. YOLOv8-n was adopted as the baseline, with Ghost conv applied to the backbone for compression. Lightweight channel-wise attention was introduced in the neck and the head via SE blocks to enhance feature representation without sacrificing real-time performance. By comparing the detection precision and speed in CPU and four GPUs with different performances, this study explores the feasibility of applying a deep learning-based gastric cancer detector in processors in actual medical field. Experimental results demonstrate that the proposed model maintains real-time inference speed on GPUs of various performance levels. Moreover, it achieved 77.5 mean Average Precision (mAP) which outperformed the mAP of 76.5 of YOLOv8-m, while outperforming 74.4 mAP of YOLOv8-n (baseline) by 4.16%. The complexity of the proposed model was minimal, with only 2.8M parameters and 7.7 GFLOPs, demonstrating the achievement of high detection precision with a reduced number of parameters.