This paper addresses option pricing by introducing a 3D tensor-based data framework that integrates three information sources: fundamental data (e.g., spot price, strike price), price data (e.g., previous settle price), and Greeks (e.g., Delta, Gamma). The data is organized as a 3D tensor with three channels, each containing a 2D matrix of options versus features over time. Two deep learning models are proposed: a CNN+RNN model that uses 1D convolutions for spatial feature extraction followed by GRU layers for temporal modeling, and a Conv-LSTM model that adapts the standard convolutional LSTM by replacing 2D convolutions with 1D convolutions to match the tensor shape. Experiments on Chinese 50ETF option data from April 2018 to June 2020 show that both models outperform the Black-Scholes formula and a standard LSTM in terms of MSE, RMSE, and MAP. The Conv-LSTM with three channels achieves the best performance (MSE=0.00019, RMSE=0.01396), demonstrating the effectiveness of the tensor representation and the modified Conv-LSTM architecture. However, the Black-Scholes model yields lower MAPE, suggesting room for improvement in deep learning models for low-price scenarios. The study highlights the potential of tensor-based deep learning for financial time series prediction.