Objective:
The aim of this study is to improve the segmentation and classification of lung regions from CT images of 78 chest cancer patients, focusing on enhancing the accuracy in detecting cancerous tissues. This research evaluates the performance of different U-Net backbone models (VGG16, ResNet50, Xception) in segmentation and employs a novel BIR-enhanced CNN model for classifying lung injury severity.
Methodology:
A U-Net model with three different backbones—VGG16, ResNet50, and Xception—was utilized for lung region segmentation. Preprocessing techniques such as CLAHE (Contrast Limited Adaptive Histogram Equalization) were applied to enhance contrast and image quality, followed by resizing to 128x128 pixels and normalization. For classification, a BIR-enhanced CNN model was employed to assess lung injury severity. The models were evaluated across multiple metrics, including accuracy, recall, F1 score, Intersection over Union (IoU), and Dice coefficient.
Results:
Among the models, VGG16 achieved the highest performance, with an accuracy of 0.9836 ± 0.0177, recall of 0.9696 ± 0.0737, F1 score of 0.9363 ± 0.0832, IoU of 0.8893 ± 0.1178, and Dice coefficient of 0.9363 ± 0.0832 in segmentation tasks. For the classification of lung injury severity, the BIR-enhanced CNN model, also utilizing VGG16, achieved a classification accuracy of 97.83%.
Conclusion:
This study demonstrates the significant impact of preprocessing on segmentation and classification performance. The U-Net model with the VGG16 backbone not only provides highly accurate segmentation of lung regions but also highlights cancerous areas effectively. The integration of the BIR model further improves classification accuracy, indicating that this combination offers an effective approach for lung cancer detection and diagnosis.