Automated image semantic segmentation plays a crucial role in computer-aided disease diagnosis. One of the most popular models used in this field is UNet, known for its structural efficiency and adaptability to various datasets. While many UNet architectures exist for medical applications, none have been developed independently without using any existing packages. The primary goal of this research is to investigate the design parameters in developing a U-Net architecture with data pre-processing and a feeding mechanism. It also aims to understand the working principle of the developed model and evaluate its performance parameters without any involvement of existing packages such as TensorFlow and PyTorch. This forward-backward path design model utilizes two publicly available datasets, specifically for cell nuclei and brain tumors. It effectively learns through backpropagation, reducing loss by up to 2% and achieving a maximum Dice similarity coefficient of 91%. Furthermore, this UNet model offers access to a complete UNet library built from scratch, without compromising performance metrics. This work signifies a valuable contribution to performance-aware design, custom hardware acceleration, and a comprehensive theoretical understanding of elements that are essential for research, optimization, and real-world application.