Hybrid Curriculum Learning for Data-Efficient Lung Nodule Detection with YOLOv11

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate pulmonary nodule detection on chest CT is vital for lung cancer screening.
• High annotation costs hinder the development of robust nodule detectors.
• This study addresses data efficiency in training these detectors.

Purpose of the Study:

• To systematically study curriculum learning for CT-based lung nodule detection.
• To propose a hybrid curriculum learning framework for data-efficient optimization.
• To evaluate the framework on the enhanced LUNA25 benchmark.

Main Methods:

• A hybrid curriculum learning framework was developed, fusing handcrafted features (nodule size, count) and model-driven signals (prediction confidence).
• A three-stage curriculum progressed from easy to hard samples.
• The approach used YOLOv11s as a baseline and compared it to conventional training and single-source curricula.

Main Results:

• Hybrid curriculum learning improved mAP50 (0.672 to 0.696) and mAP50-95 (0.369 to 0.385) on the LUNA25 test set.
• Recall increased from 0.588 to 0.634, and precision from 0.725 to 0.764.
• Consistent gains were observed in data-efficiency experiments with reduced datasets.

Conclusions:

• Combining image complexity and optimization feedback offers effective sample scheduling for lung nodule detection.
• The proposed hybrid curriculum learning framework enhances robustness and data efficiency.
• This approach is promising for optimizing lung cancer screening tools.