A neighbor pixel modelled ROI for mammogram classification

The effective selection of the Region of Interest (ROI) plays a vital role in enhancing the performance of computer-aided detection and diagnosis (CADx) systems. Typically, ROIs are either defined by radiologists or obtained after segmentation for feature extraction. Recent CADx systems utilizing neural networks often process the entire mammogram, resulting in increased computational cost and reduced accuracy, especially when the mass occupies only a small portion of the image. To address this, researchers have introduced various ROI selection methods that eliminate non-mass regions, aiming for better diagnostic precision. Among these, the Sparse-ROI technique stands out by modeling irregular mass shapes using sparse matrices. While this approach significantly reduces computation time for mass characterization via statistical matrices, it struggles with Gray Level Aura Matrix (GLAM) construction in wider ROIs. To overcome this limitation, we propose an enhanced model, the Neighbour Pixel Modelled ROI (NPMR), which incorporates four neighboring pixels to improve GLAM efficiency. Experimental evaluation using the MIAS database (322 mammograms) shows that NPMR performs well in both classification and GLAM construction. Comparative analysis with the traditional Fixed Window and Sparse-ROI methods demonstrates that NPMR achieves remarkable computational efficiency—reducing processing time by 99.8% and 99.9%, respectively—while maintaining a high classification accuracy of 97.2%. Metrics such as accuracy (97.2%), precision (97.82%), sensitivity (74.48%), size (94.8%-pixel reduction), and execution time confirm the superiority of the proposed NPMR approach in mammographic analysis.