A Novel Deep Learning Model Based Cancerous Lung Nodules Severity Grading Framework Using CT Images

Abstract

Lung cancer remains one of the leading causes of cancer-related mortality, with early diagnosis being critical for improving patient survival rates. Existing deep learning models for lung nodule severity classification face significant challenges, including overfitting, computational inefficiency, and inaccurate segmentation of nodules from CT images. To overcome these limitations, this study proposes a novel deep learning framework integrating a Quadrangle Attention-based U-shaped Convolutional Transformer (QA-UCT) for segmentation and a Spatial Attention-based Multi-Scale Convolution Network (SMCN) for classification. CT images are enhanced using the Rotationally Invariant Block Matching-based Non-Local Means (RIB-NLM) filter to remove noise while preserving structural details. The QA-UCT model leverages transformer-based global attention mechanisms combined with convolutional layers to segment lung nodules with high precision. The SMCN classifier employs spatial attention mechanisms to categorize nodules as solid, part-solid, or non-solid based on severity. The proposed model was evaluated on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset. This proposed model achieves a 98.73% dice score for segmentation and 99.56% classification accuracy, outperforming existing methods such as U-Net, VGG, and autoencoders. Improved precision and recall demonstrate superior performance in lung nodule grading. This study introduces a transformer-enhanced segmentation and spatial attention based classification framework that significantly improves lung nodule detection accuracy. The integration of QA-UCT and SMCN enhances both segmentation precision and classification reliability. Future research will explore adapting this framework for liver and kidney segmentation, as well as optimizing computational efficiency for real-time clinical deployment.