Automatic Cluster Selection in K-Means Lung Segmentation

Lung segmentation is a critical step in machine-learning-based radiomics using thoracic computed images. It involves isolating a specific region of interest, but variations in lung intensity values caused by diseased lung tissue can difficult correct segmentation. Although K-means is commonly used, it requires manual intervention to select each cluster related to the region of interest, leading to an efficiency decrease in terms of the specialist’s time and effort, especially for large image volumes. To address these limitations, an automatic cluster selection methodology is proposed. It involves a training process to determine a threshold for discriminate clusters; then, morphological transformations and image processing techniques enhance segmentation. Evaluation using DICOM images from the Interstitial Lung Diseases Database yielded a Jaccard Similarity Index of 0.9056 and a Dice Similarity Coefficient of 0.9475, demonstrating the effectiveness and accuracy of the proposed approach.