An Automated Diagnosis Method for Lung Cancer Target Detection and Subtype Classification-Based CT Scans

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer Vision for Healthcare

Background:

• The YOLO V8 algorithm faces challenges with false positives and missed detections when identifying small lung cancer targets.
• Accurate detection of small pulmonary nodules is critical for early lung cancer diagnosis and treatment planning.

Purpose of the Study:

• To develop an enhanced YOLO V8 detection model specifically for improving the accuracy of small lung cancer target detection.
• To address limitations in existing models regarding feature representation, robustness, and small object localization.

Main Methods:

• Integration of a large separable kernel attention mechanism into the C2f module to broaden information retrieval.
• Strengthening feature extraction in the Backbone and enhancing multi-scale feature interaction in the Neck section.
• Embedding depth-wise convolution and Coordinate Attention in the Fast Spatial Pyramid Pooling module, and introducing a Minimum Point Distance-based IOU loss.

Main Results:

• The enhanced YOLO V8 model demonstrated superior performance compared to mainstream detection networks, achieving higher average precision values.
• The improved network surpassed other classification networks in accuracy for lung cancer auxiliary diagnosis.
• Experimental validation confirmed enhanced feature representation, robustness, and reduced feature loss, leading to improved detection accuracy for small targets.

Conclusions:

• The proposed enhanced YOLO V8 model significantly improves the localization and recognition of small lung cancer targets.
• The model's architectural modifications and novel loss function contribute to its outstanding performance in auxiliary lung cancer diagnosis.
• This research offers a promising advancement in AI-driven tools for early and accurate lung cancer detection.