With the significant increase in the use of deep learning (DL) for biomedical imaging, the corresponding DL models have become increasingly complex and computationally intensive to achieve high accuracy. This work presents both architecture-aware optimizations and sparsity optimizations to efficiently utilize underlying parallel hardware resources and reduce the computational demand of DL models while maintaining their accuracy. We demonstrate the efficacy of our optimization techniques on an existing DL model in the biomedical domain, i.e., DDNet, short for Densenet and Deconvolution Network, that is designed to enhance the quality of CT images. Overall, our optimization techniques in concert reduce the total training time by 1.94 \(\times \) while maintaining accuracy.