CNN-Based Classification and Segmentation of Pancreatic Tumor Detection

Abstract

The tumors of the pancreas are linked with high mortality rates due to late discovery of the diseases, showing the need for timely and accurate diagnosis. The fire deep learning architecture, proposed in this research, has been shown to be a hybrid system for categorizing pancreatic tumors from computed tomography (CT) images. This involves training a U-Net to segment images and applying Convolutional Neural Networks (CNNs) to categorize computed tomography (CT) images. It is performed using a publicly available dataset of female pancreatic CT scans. Preprocessing was also performed for model generalization through normalization, resizing, and data augmentation. U-Net was used to perform pixel-wise segmentation of the tumor, yielding a Dice Coefficient of 0.89, i.e., very high segmentation accuracy. These results were then fed into a CNN classifier, which achieved 93 percent accuracy in determining whether a tumor is present. The results of the model training and testing procedure are presented. The effectiveness of the models is measured using quantitative performance metrics, including the Dice Coefficient, classification accuracy, precision, recall, F1-score, ROC curves, and confusion matrices. Accordingly, it has been executed by comparing a Python environment on Google Colab, using the PyTorch deep learning library, for pancreatic tumor detection.