Utilising deep learning for classification of disease-related lung opacities through colourmap optimisation

Introduction: Existing deep learning models for lung opacity detection primarily focus on grayscale images, overlooking the potential benefits of colour map transformations. In this study, we address this gap by fine-tuning DarkNet-53 and ResNet-101 deep learning models using both grayscale and 16 distinct colour maps to enhance disease classification. Through transfer learning and data augmentation, we explore how colour map transformations can improve the models’ diagnostic accuracy, sensitivity, and specificity, aiming to optimize deep learning performance for better clinical decision-making. Methods: A total of 11,342 chest X-ray images, consisting of normal and disease-related lung opacity images, were used in this study. These images were pre-processed into 16 different colourmaps to enhance visualization. The DarkNet-53 and ResNet-101 deep learning models were fine-tuned using transfer learning and standard data augmentation techniques. Performance metrics, including accuracy, sensitivity, and specificity, were calculated to evaluate the models. Results: The DarkNet-53 model achieved an average accuracy of 89.9%, sensitivity of 88.1%, and specificity of 92.9% across various colourmaps. The ResNet-101 model demonstrated similar performance with an average accuracy of 89.9%, sensitivity of 88.2%, and specificity of 92.8%. The “Spring” colourmap resulted in the highest accuracy and sensitivity for the DarkNet-53 model, while the “Copper” colourmap was optimal for the ResNet-101 model. Conclusion: The findings highlight the importance of optimizing colourmap selection to enhance the performance of deep learning models for disease-related lung opacity detection. The careful choice of colourmaps can significantly improve model accuracy, sensitivity, and specificity, leading to better diagnostic precision and patient outcomes in managing respiratory diseases. Malaysian Journal of Medicine and Health Sciences (2024) 20(SUPP10): 1-9. doi:10.47836/mjmhs.20.s10.1