Medical Image Segmentation for Anomaly Detection using Deep Learning Techniques

Medical images are the standard approach for the analysis and diagnosis of critical issues of diseases. To minimize the time-consuming inspection and evaluation process of the medical images from physicians in diagnosis, an automatic segmentation mechanism of abnormal features in medical images is required. To address the limited availability of medical image data, deep learning frameworks for multi-class image segmentation have been implemented. Moreover, the existing deep learning frameworks are static. Hence to make dynamic for the image segmentation purpose, the existing deep learning-based frameworks for medical image segmentation have been updated by integrating advanced architectures to observe performance enhancements. The existing UNet model has been integrated with a vision transformer to capture the structural properties of the medical image. Similarly, the ResNet50 architecture has been integrated with DeepLabv3plus for better extraction of features. The CVC-ClinicDB dataset, ISIC dataset, Brain Tumor dataset and HyperKVasir dataset have been collected from multiple sources. The collected datasets have been processed with image augmentation, Contrast Limited Adaptive Histogram equalization (CLAHE), and normalization. The preprocessed image dataset has been categorized into training, validation, and testing parts and used accordingly. The training image data has been used to train the multiple deep-learning models. The adaptive moment (Adam) optimizer has been used for the optimization process. The loss measurement has been carried out using categorical cross-entropy. The trained models have been evaluated using a testing dataset. The performance of the implemented deep learning framework has been measured by Accuracy, Precision, Recall, F1-Score, Dice Coefficient, and Validation Dice Coefficient metrics.