RUMS-net: A Deep Learning Framework for Accurate Segmentation and Classification of Lung Cancer Images

Abstract

This study gives a precise methodology to accurately process and classify thoracic cancer images using the The Lung Image Database Consortium image collection dataset. Given  proposed approach utilizes the RUMS-net model, which combines ResNet-50, U-Net, and Multi-Class SVM algorithms. The process commences by resizing images utilizing bicubic interpolation, subsequently applying bilateral filtering to diminish noise while maintaining edges. Contrast enhancement is accomplished through the use of Bi-Histogram Equalization with Brightness Preservation Technique (BBHE), and the input image is then transformed to a binary format using an iterative selection thresholding method. The binary image is refined through morphological opening operations using disk-shaped structuring elements. Subsequently, the refined image is inverted and utilized for active contour segmentation employing the Random Walker algorithm. The inverted images are overlaid with segmented contours, and then Wiener filtering is implemented. The image undergoes a conversion to binary, followed by hole filling. Circle detection is then used to identify regions of interest, which are overlaid onto the original image. The segmented image is obtained through the subtraction of binary images, and quantitative metrics such as the total area of the tumor and the ratio of affected area are calculated. The RUMS-net model utilizes ResNet-50 for feature extraction, U-Net for segmentation, and Multi-Class SVM for classification in order to perform classification. The proposed model attains 99.75% as accuracy rate, sensitivity of 99.91%, 98.23% specificity, and precision of 99.79%, surpassing other models and showcasing its effectiveness in delivering accurate and dependable medical diagnosis and analysis.