This research article discusses the implementation aspects of a Deep Learning architecture based on U-Net for medical image segmentation. A base model of the U-Net architecture is extended and experimented. Unlike the existing model, the input images are enhanced by applying a Non-Local Means filter optimized using a metaheuristic Grey wolf optimization method. Further, the model parameters are modified to achieve better performance. Tests were performed using two benchmark B-mode Ultrasound image datasets of 200 Breast lesion images and 504 Skeletal images. Experimental results demonstrate that the modifications resulted in more accurate segmentation. The performance of the modified implementation is compared with the base model and a Bidirectional Convolutional LSTM architecture.