CRCFusionAICADx: Integrative CNN-LSTM Approach for Accurate Colorectal Cancer Diagnosis in Colonoscopy Images

Colorectal cancer (CRC) is a critical health issue worldwide and is very treatable if diagnosed on time. This paper proposes an innovative CADx system, namely CRCFusionAICADx, that enhances the efficiency of diagnosis by fusing CNNs with LSTM networks and feature integration techniques. Using data from the CKHK-22 colonoscopy image dataset, we preprocess the images into grayscale first and then apply LBP analysis for emphasizing textural features. These are further analyzed using three different pre-trained CNN models: VGG16, DenseNet-201, and ResNet50. These were chosen because of their complementary feature extraction capabilities. The resultant features from grayscale, LBP, and raw images will be fused to create an integrated dataset. To increase variability in the dataset and reduce overfitting for the network, we decided to apply a series of data augmentation techniques, which included zooming in, rotation, and horizontal flipping. By doing so, we expanded the dataset into 57,148 images. This augmented dataset is then used to train a model, RDV-22, which includes an integration of the architectures of VGG16, DenseNet-201, and ResNet50, with CNN and CNN + LSTM layers. The LSTM network learns the temporal dependencies of frames in a sequence and hence allows for more sensitive and specific detection of CRC. CRCFusionAICADx produces very impressive results, where the RDV-22 model produces a testing accuracy of 90.81%, precision of 91.00%, recall of 90.00%, and an F1 score of 90.49% in its results. This gives the model an ROC AUC of 0.98, reflecting very strong discriminatory power. This integrative approach thus shows tremendous promise for early CRC detection by offering a strong diagnostic tool that integrates both spatial and temporal features into a new standard in clinical diagnostics.