Lung cancer detection using image processing: a hybrid CNN-DBN framework for accurate and efficient classification

Abstract

The study introduces a straightforward but powerful approach that blends today’s deep-learning tools with tried-and-true image-processing tricks for spotting lung cancer on CT scans. At its heart sits CLAHE-Contrast Limited Adaptive Histogram Equalization-a workhorse in medical imaging that sharpens contrast just where small shifts in tissue density matter most for diagnosis. After that, the enhanced images are fed into a custom-built Convolutional Neural Network (CNN) that has the ability to learn strong and discriminative features with a compact architecture. These are then fed through a Deep Belief Network (DBN), made up of stacked layers of Restricted Boltzmann Machines (RBMs) that can learn hierarchical representations in an unsupervised way. These ultimate representations are then fed through classification via logistic regression, a straightforward yet effective supervised learning technique which can take advantage of the good quality of the learned embeddings. The model avoids overfitting by decoupling feature learning and decision making and employs a refinement mechanism that identifies and reprocesses edge-case test samples through the CNN and DBN recurrently for improved prediction stability. With a test accuracy of 98.63% and a FLOP count of merely 47.5 million, this model achieves a remarkable trade-off between accuracy and computational resource usage. Being highly scalable and modular, it is particularly suited for deployment in clinical settings, specifically on edge devices with limited resources. This pipeline provides a good foundation for pursuing cost-effective, high-accuracy solutions for medical imaging tasks in future studies.