Enhancing Colorectal Cancer Diagnosis With Feature Fusion and Convolutional Neural Networks

Abstract

TumorDiagX is a cutting‐edge framework that combines deep learning and computer vision to accurately identify and classify cancers. Our collection of colonoscopies 1518 images is meticulously pre‐processed, including greyscale conversion and local binary pattern (LBP) extraction, before being securely stored on the Google Cloud platform. In the second phase, we fully assess three different convolutional neural networks (CNNs): residual network with 50 layers (ResNet‐50), DenseNet‐201 and visual geometry group with 16 layers (VGG‐16). Stage three introduces four integrated CNNs (ResNet‐50+DenseNet‐201 (RD‐22), DenseNet‐201+VGG‐16 (DV‐22), ResNet‐50+VGG‐16 (RV‐22), and ResNet‐50+DenseNet‐201=VGG‐16 (RDV‐22)) to improve cancer detection by combining the capabilities of several networks. Comprehensive analysis and training on the datasets provide significant insights into CNN’s performance. The fourth step involves an extensive comparison, integrating and comparing all three data sets using individual and integrated CNNs to determine the best effective models for cancer diagnosis. In this final step, image segmentation leverages an encoder–decoder network, namely a Universal Network (U‐Net) CNN, to aid in the visual detection of malignant cancer lesions. The results highlight the effectiveness of TumorDiagX, with the feature fusion CNN using DenseNet‐201 attaining training and testing accuracies of 97.27% and 97.35%. Notably, CNN (feature fusion) in combination with RDV‐22 performs better, with training and testing accuracy of 98.47% and 97.93%, respectively, and a dice coefficient of 0.92. The information is privately maintained in the cloud and acts as an essential asset for healthcare practitioners, allowing for specific cancer prediction and prompt detection. Our method, with its meticulous performance metrics and multifaceted approach, has the potential to advance early cancer identification and treatment.