Feature stabilization in convolutional neural networks using Proportional Integral Controller for lung nodule...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Control Theory

Background:

• Accurate lung nodule classification from CT images is challenging due to image variations and unstable deep learning features.
• Convolutional Neural Networks (CNNs) show promise but struggle with convergence and generalization, especially with limited data.

Purpose of the Study:

• To develop a control-inspired CNN framework for stable and reliable lung nodule classification.
• To regulate feature representations during deep network training using a Proportional Integral Controller (PIC).

Main Methods:

• Proposed a CNN framework integrating a PIC as a preprocessing module and an intermediate layer.
• Investigated both manually tuned and automatically learned PIC configurations.
• Utilized proportional and integral components to manage feature deviations and accumulated errors.

Main Results:

• Achieved state-of-the-art performance on the IQ-OTH/NCCD lung cancer dataset (Accuracy 0.96, F1-score 0.96).
• Eliminated false positives with a Precision of 0.93.
• Demonstrated strong generalization through cross-dataset validation and confirmed the importance of PIC integration.

Conclusions:

• Integrating control theory principles into CNNs offers an effective strategy for enhancing medical image classification.
• The proposed PIC-regulated framework improves stability, generalization, and interpretability in lung nodule detection.
• This approach addresses key challenges in deep learning for medical image analysis.