Evaluation of the Hybrid Model of K-NN and Bi-ConvLSTM in Identifying Lung Cancer Nodules in CT Images: An Optimal Study

Abstract

Purpose: The objective of this study is to develop a hybrid model combining K-Nearest Neighbor (K-NN) and Bidirectional Convolutional Long Short-Term Memory (BiConvLSTM) for effective detection of lung cancer nodules in chest CT images.

Materials and Methods: A dataset of 200 high-resolution CT images (100 nodular, 100 healthy controls) was utilized. Preprocessing steps, including wavelet-based noise reduction and GAN-enhanced contrast refinement, ensured standardized image quality. Spatial features were initially extracted using K-NN, optimized across multiple values of k. Temporal dependencies were processed through BiConvLSTM, supported by residual pathways to preserve information integrity. Evaluation metrics and advanced optimizers (Adam, Nadam, and AdamW) were compared to assess performance.

Results: The hybrid model achieved peak accuracy (97.2% ± 1.1) with k = 7, while k-fold cross-validation with k = 5 maintained strong performance. AdamW optimizer outperformed others in generalization metrics and Dice similarity (0.93), while Nadam demonstrated faster convergence, reducing training epochs by 20%. Metrics such as AUROC (0.98) and the Jaccard index confirmed the robustness of the model compared to conventional architectures like U-Net. Additionally, the statistical significance of the proposed model was validated with a P-value of 1.2E-5.

Conclusion: This proposed hybrid AI model addresses challenges in nodular identification and provides a scalable, reliable solution for lung cancer detection in clinical imaging workflows.