Lung cancer is one of the most common causes of death from cancer around the world, and catching it early through computed tomography (CT) scans can drastically improve survival. However, automated classification of pulmonary nodule candidates is hard because images do not all have the same intensity across scanners and protocols, resulting in inconsistent performance, more false positives (FP), and a ceiling on how much deep learning models work in an average clinic. In this work, we tackle this by introducing a preprocessing step that corrects intensity differences before feeding images into classification models. We use Contrast-Limited Adaptive Histogram Equalization (CLAHE), but with its key parameters tuned automatically via a modified version of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This helps to boost local contrast adaptively, keeps important anatomical details intact, and cuts down on noise. We tested the approach on the public LUNA16 dataset, first checking image quality (Peak Signal-to-Noise Ratio (PSNR) around 53 dB and Structural Similarity Index (SSIM) of 0.9, better than standard methods), then training three popular deep models—namely, ResNet-50, EfficientNet-B0, and InceptionV3—with CutMix augmentation for better generalization. On the enhanced images, ResNet-50 achieved up to 99.0% classification accuracy with substantially less FP than when using the raw scans. Taken together, these results demonstrate that intelligent and optimized preprocessing can effectively mitigate intensity variations via deep learning for lung nodule detection, thus coming closer to realizing the practical toolbox of computer-aided diagnosis in routine clinical practice.
Khan et al. (Sun,) studied this question.