A lightweight YOLOv8-based model for gastric cancer detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deep learning models show promise for gastric cancer detection, but their performance is processor-dependent.
• Real-world medical environments utilize diverse computer processors (CPUs and GPUs) with varying performance levels.

Purpose of the Study:

• To develop a gastric cancer detection model using YOLOv8 that achieves real-time performance and is less sensitive to processor variations.
• To evaluate the model's feasibility in actual medical fields across different hardware.

Main Methods:

• Proposed a modified YOLOv8 model incorporating Ghost convolution in the backbone and SE blocks for channel-wise attention in the neck and head.
• Evaluated the model's detection precision and inference speed on a CPU and four GPUs with varying performance.
• Compared the proposed model against baseline YOLOv8-n and YOLOv8-m.

Main Results:

• The proposed model achieved 77.5% mean Average Precision (mAP), outperforming YOLOv8-m (76.5% mAP) and YOLOv8-n (74.4% mAP) by 4.16%.
• Maintained real-time inference speed across different GPU performances.
• The model has minimal complexity with 2.8M parameters and 7.7 GFLOPs.

Conclusions:

• The developed YOLOv8-based model offers high detection precision and real-time performance, adaptable to various computer processors.
• This research supports the practical application of deep learning for gastric cancer detection in diverse clinical environments.