Fine-grained lung cancer object detection via dilated reparameterization and explicit positional gating optimization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early lung cancer detection and classification are vital for treatment planning.
• Current computer-aided diagnosis (CAD) systems often perform binary classification or separate feature classification and staging.
• Technical challenges include scale variation, subtle inter-class differences, and complex backgrounds.

Purpose of the Study:

• To propose an end-to-end fine-grained lung cancer object detection algorithm using an improved YOLO architecture.
• To address limitations in existing CAD systems for complex lung cancer classification tasks.
• To achieve unified detection and classification of normal tissue, non-small cell lung cancer (NSCLC) subtypes, and TNM stages.

Main Methods:

• Developed an improved YOLO architecture for a 7-class detection task (normal, 3 NSCLC subtypes, anatomical locations with TNM stages).
• Incorporated a Dilated Reparameterization Block (C3k2_DRB) for enhanced multi-scale feature extraction.
• Introduced an Explicit Positional Gating Optimization (C2PSA_EPGO) attention mechanism for adaptive focus on lesion textures.
• Employed a dynamic upsampling strategy (DySample) for content-aware spatial alignment and boundary preservation.

Main Results:

• Achieved 97.6% Precision and 99.2% mean Average Precision (mAP@0.5) on a dataset of 1886 images.
• Maintained low parameters (2.46 M) and computational cost (6.3 GFLOPs).
• Demonstrated superior localization precision and fine-grained classification compared to mainstream object detection algorithms.

Conclusions:

• The proposed algorithm offers a unified framework for complex, fine-grained lung cancer detection and classification.
• Architectural enhancements significantly improve performance in medical vision tasks.
• The method shows potential for advancing early diagnosis and clinical treatment planning in lung cancer.