Recent advancements in theory and empirical studies have demonstrated the powerful capabilities of neural networks in finance. In this work, we investigate the option pricing problem in discrete time using neural networks, focusing on experiments conducted with the discretized Heston model for a call option. Specifically, we utilize Long Short-Term Memory (LSTM) networks to address both the super-hedging and quantile hedging problems. Our numerical results demonstrate that the option prices obtained using the LSTM-based approach perform well within the hedging framework under the Heston model. Notably, the method remains effective even when transaction costs are incorporated. This adaptability to markets with frictions highlights a novel contribution of our work, as traditional hedging strategies often struggle in the presence of transaction costs.