In this chapter, we present a lightweight and secure neural network (NN) inference service framework, MediSC. MediSC is specifically designed for medical diagnostic services, enabling enterprises to provide secure medical diagnostic capabilities to their customers by executing NN inference in the ciphertext domain. MediSC guarantees the privacy of both parties through robust cryptographic techniques. At its core, we introduce an efficient and communication-optimized secure inference protocol. Our protocol relies exclusively on lightweight secret-sharing techniques and is capable of handling commonly-used linear and non-linear NN layers effectively. In comparison to solutions based on garbled circuits, MediSC achieves significantly better performance, with 24 \(\times \) lower latency and 868 \(\times \) less communication for secure ReLU, and 20 \(\times \) lower latency and 314 \(\times \) less communication for secure Max-pool. We evaluate MediSC using two benchmark datasets and four real-world medical datasets, performing a comprehensive comparison with prior works. The results highlight the exceptional performance of MediSC, demonstrating that it is substantially more bandwidth-efficient than existing approaches.