Deep learning driven lung cancer prediction using modified u-net segmentation and CNN classification

Abstract

Lung cancer remains a major contributor to global cancer mortality, highlighting the urgent need for accurate and efficient early diagnostic tools. This study proposes a deep learning–based computer-aided diagnosis (CAD) framework for automated lung nodule segmentation and malignancy classification using the publicly available LIDC-IDRI CT dataset. The framework combines a modified U-Net segmentation network a residual (ResNet-34) encoder, attention gates at the skip connections, and an Atrous Spatial Pyramid Pooling (ASPP) module for multi-scale feature fusion with a custom Squeeze-and-Excitation (SE) CNN classifier, trained using a two-stage optimisation strategy. On a held-out, patient-level test split, the segmentation network achieved a Dice score of 0.847 (IoU 0.736), and the classification network achieved a mean accuracy of 94.5% (± 0.3% SD across five training runs; 95% CI ≈ 94.2–94.8%), with sensitivity of 92.3%, specificity of 96.1%, and AUC-ROC of 0.967, outperforming ResNet-50, VGG-16, and DenseNet-121 baselines trained under an identical protocol. These results are competitive with, and in several respects exceed, recently reported approaches on the same dataset, although a transformer-based segmentation model achieved a marginally higher Dice score at substantially greater computational cost. The framework was evaluated on a single, retrospective, single-institution dataset; prospective, multi-institutional validation, together with an assessment of end-to-end deployment latency, is required before any claim of clinical readiness can be made. These findings indicate the potential of the proposed CAD framework to support radiologists as a second-reader tool in lung cancer screening, pending further external validation.