Abstract
The presentation of colorectal diseases based on deep learning has attracted much research since it offers the development of correct and machine-driven diagnosis. The current study advents a new DenseNet-121 model with Squeeze-and-Excitation (SE) blocks that can enjoy better feature extraction and classification into colon diseases. The sample has four classes and 1,500 images each. In order to enhance the robustness of the models, the preprocessing pipeline is done completely, involving resizing of images, Gaussian smoothing, and Otsu thresholding. DenseNet-121 model incorporating SE blocks is trained on an 80:20 training-testing split with data augmentation being used to deal with over-fitting. The performance assessment is performed as precision, recall, F1-score, and inference time in comparison to the classical architecture of ResNet-50 and VGG-16 with DenseNet-121. The results obtained on the experimental show that the proposed model scored 96.3 percent accuracy, 97.1 percent precision, 95.7 percent recall, and 96.4 percent F1-score with better performance on the colorectal disease category than the baseline models. Also, dimensionality is reduced because of the presence of Global Average Pooling (GAP), and the model is made more discriminative thanks to SE blocks that allow the recalibration of features. The analysis of confusion matrix proves high classification reliability indicating that there is minimal misclassification among the disease categories. The present article reveals the case of successful deep learning application in the analysis of endoscopy images and opens up the possibility of real-time application in clinical practice in the form of a computer-aided diagnosis system.
Keywords: Colorectal Disease, Computer-Aided Diagnosis System, Deep Learning, DenseNet-121, Global Average Pooling, Image Segmentation.