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

INTRODUCTION: Reliable classification of lung nodules from computed tomography (CT) images remains a challenging problem due to variations in image intensity, noise, and unstable feature representations during deep network training. Although convolutional neural networks (CNNs) have achieved promising results in medical image analysis, their internal feature dynamics are often difficult to control, which can affect convergence stability and generalization, particularly when working with limited clinical data.

METHOD: In this work, we propose a control-inspired CNN framework that incorporates a Proportional Integral Controller (PIC) to regulate feature representations during the learning process. The PIC is integrated into the network in two distinct ways: as a preprocessing module before the CNN and as an intermediate layer embedded within the convolutional architecture. Both manually tuned and automatically learned PIC configurations are investigated to analyze the influence of fixed, knowledge-driven control parameters vs. adaptive, data-driven feedback mechanisms. The proportional component responds to instantaneous feature deviations, while the integral component compensates for accumulated errors, jointly contributing to more stable and consistent feature learning.

RESULT: The proposed approach is evaluated on the IQ-OTH/NCCD lung cancer dataset using standard classification metrics. The proposed method achieves state-of-the-art performance (Accuracy 0.96, F1-score 0.96) and eliminates false positives (Precision 0.93). Ablation and statistical analyses confirm the importance of PIC placement and parameter tuning, while cross-dataset validation demonstrates strong generalization. Overall, this study demonstrates that integrating principles from control theory into deep learning architectures provides an effective and interpretable strategy for enhancing medical image classification.