Comparative Analysis of Colour Normalisation Techniques for Lung Carcinoma Classification using Whole Slide Images: A Retrospective Computational Study

Introduction: Lung cancer continues to be the top cause of cancer-related deaths worldwide, with around 2.48 million new cases and 1.8 million fatalities reported in 2022. Histopathological Whole Slide Images (WSIs) play a crucial role in diagnosing different lung cancer subtypes. However, variations in Haematoxylin and Eosin (H&E) staining and image acquisition can lead to colour inconsistencies, which in turn affect the performance of automated diagnostic models that rely on Convolutional Neural Networks (CNNs). Aim: To improve the accuracy and robustness of automated classification of lung carcinoma from whole slide histopathological images, by implementing four colour normalisation techniquesHistogram Matching, Reinhard, Macenko, and Vahadane. Materials and Methods: The present study is a retrospective in nature and was conducted in Indian Council of Medical Research (ICMR) funded Lab, Brainware University, Kolkata, India for a period of two years. The research study was conducted with a total of 5,000 annotated WSI patches (each measuring 512×512 pixels), evenly split between Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LSCC). Each normalisation technique viz., histogram matching, reinhard, macenko, and vahadane was applied to the dataset, and the processed images were evaluated using the Structural Similarity Index (SSIM), Quality SSIM (QSSIM), and CNN-based classification accuracy. All models were developed using Python, specifically the TensorFlow and OpenCV libraries, and were trained on standardised datasets. Results: Among the methods assessed, Reinhard normalisation achieved the highest classification accuracy at 91%, followed closely by Macenko and Vahadane at 88%, and histogram matching at 84%. Both Macenko and Vahadane techniques excelled in preserving tissue morphology, while Reinhard normalisation struck a good balance between performance and computational efficiency. Analyses using SSIM and QSSIM confirmed that normalisation significantly improved structural and colour consistency across WSIs. Conclusion: Colour normalisation plays a vital role in enhancing the reliability of CNN-based histopathological classification for lung carcinoma. The Reinhard method stood out as the most effective due to its impressive accuracy and efficiency, although advanced techniques like Vahadane also showed promise. However, analysis of training and validation performance revealed a substantial gap across all techniques, indicating overfitting and limited generalisation.