Improving Deep Learning Based Lung Nodule Classification Through Optimized Adaptive Intensity Correction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Lung cancer diagnosis relies heavily on early detection via CT scans.
• Inconsistent image intensity across scanners hinders automated pulmonary nodule classification accuracy.
• Deep learning models face performance limitations due to image variations and false positives.

Purpose of the Study:

• To develop and evaluate an optimized image preprocessing technique for lung nodule detection.
• To address intensity variations in CT scans for improved deep learning model performance.
• To reduce false positives in automated lung nodule classification for clinical utility.

Main Methods:

• Implemented Contrast-Limited Adaptive Histogram Equalization (CLAHE) with automated parameter tuning using Covariance Matrix Adaptation Evolution Strategy (CMA-ES).
• Applied preprocessing to the LUNA16 dataset, assessing image quality using PSNR and SSIM.
• Trained deep learning models (ResNet-50, EfficientNet-B0, InceptionV3) with CutMix augmentation on preprocessed images.

Main Results:

• CLAHE preprocessing significantly improved image quality (PSNR ~53 dB, SSIM 0.9) compared to standard methods.
• ResNet-50 achieved up to 99.0% classification accuracy on enhanced images.
• The optimized preprocessing substantially reduced false positives compared to using raw CT scans.

Conclusions:

• Intelligent and optimized preprocessing effectively mitigates intensity variations in CT scans.
• This approach enhances deep learning model performance for lung nodule detection.
• The method advances the practical application of computer-aided diagnosis in routine clinical practice.