Deep Learning for Lung Cancer Staging: A Performance Evaluation of Gamma Correction vs Histogram Equalization Using EfficientNetB0

Abstract

An accurate classification of lung cancer stages is critical in supporting the appropriate clinical decision-making and treatment strategies. This study aimed to analyze and compare the performance of two image preprocessing techniques, namely Gamma Correction and Histogram Equalization, in improving the accuracy of lung cancer stage classification based on CT scan images using the EfficientNetB0 deep learning architecture. Both methods were implemented with identical training configurations, including the use of RMSprop optimizers, last 50-layer fine-tuning, and class weighting to address data imbalances. The experimental results showed that Gamma Correction performed better with a test accuracy of 95.23% and a loss of 0.1303, compared to Histogram Equalization, which achieved an accuracy of 94.45% and a loss of 0.1523. In addition, Gamma Correction showed excellence in macro-mean F1-score metrics, especially in improving detection sensitivity in Stage Ib Adenocarcinoma and Squamous Cell Carcinoma IIIa. The training curve shows a consistent convergence trend and no indication of overfitting, with Gamma Correction demonstrating better validation stability. The results of this study confirm that the selection of preprocessing techniques has a considerable influence on the efficacy of the lung cancer stage classification model. Gamma Correction is shown to be more effective in sharpening important morphological features in CT images, while maintaining a balance between increased contrast and noise control. These findings are an important foundation for the development of accurate and reliable CAD systems for automatic lung cancer staging.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint