Abstract
Global cancer statistics indicate colon cancer caused nearly 1 million deaths, with lung cancer accounting for approximately 2 million fatalities(1). Accurate tumor identification is a critical diagnostic challenge, where histopathological examination serves as the gold standard. While current pathological localization techniques are reliable, they possess procedural limitations. This study focuses on nuclear detection and classification via pathological imaging to determine tumor presence and characterize behavior. We introduce Cross-Over LBP (CO-LBP), an innovative Local Binary Pattern variant for colon cancer diagnosis, and conduct a comparative analysis with the established step-LBP (n-LBP) method. Our evaluation framework incorporates machine learning algorithms and transfer learning techniques applied to histopathological images from the LC25000 dataset. Results demonstrate CO-LBP's clinically competitive performance (94.57% accuracy, 90.91% kappa), while n-LBP yields superior diagnostic outcomes (96.87% accuracy, 93.74% kappa), highlighting their complementary strengths in texture-based feature extraction. The experimental protocol first evaluated both LBP variants with machine learning classifiers, then with transfer learning models. The LC25000 dataset, comprising colon tissue histopathological images, served as the benchmark. Quantitative analysis revealed the following performance for n-LBP: accuracy (96.87%), kappa (93.74%), precision (96.9%), recall (96.9%), F1 score (96.9%), and ROC (99.4%). CO-LBP achieved comparable. Results accuracy (94.57%), kappa (90.91%), precision (94.9%), recall (94.9%), F1 score (94.9%), and ROC (98.8%). These findings substantiate the diagnostic potential of both methods, illustrating their distinct performance characteristics.