Fluorescence lifetime imaging technology utilizes the decay difference of the emitted fluorescences to distinguish different fluorescent molecules, which is widely used in biomedicine, chemical analysis, and life science. The quality of fluorescence lifetime imaging depends on fluorescence lifetime measurement techniques and retrieval algorithms. The measurement method based on time-correlated single-photon counting technology has become the main stream fluorescence lifetime measurement method in the field of biological research because of its high accuracy and easy low-light detection. Traditional fluorescence lifetime retrieval algorithms based on time-correlated single-photon counting technology are not suitable for the extraction of fast, high-precision, and long fluorescence lifetime. Most of the long fluorescence lifetime substances are quantum dots. In recent years, emerging deep learning techniques have also been gradually used for fluorescence lifetime retrieval, mainly realizing fluorescence lifetime imaging with fluorescence lifetimes in the range of 10 ns. Therefore, it is urgent to develop a new fluorescence lifetime retrieval algorithm to solve the constraints of retrieval accuracy and speed in a wide fluorescence lifetime range.To solve the problem of low accuracy of fluorescence lifetime retrieval in a large dynamic range, a fluorescence lifetime retrieval algorithm based on long short-term memory neural network is proposed in this paper. The algorithm uses a multi-layer long-short-term memory neural network with a time-series memory function to realize the feature extraction of the fluorescence lifetime decay histogram data which is based on time-correlated single photon counting. The unique gate structure of long-short-term memory neural network can realize the protection and control of time series information. What's more, deep learning technology is used to learn a large number of various fluorescence lifetime decay information, establish a corresponding relationship between histogram and fluorescence lifetime, and then the weight value and bias coefficients of the network are updated to make the training model more suitable for fluorescence lifetime retrieval. To train the model, the grid search method is used to select the hyperparameters of the neural network model, including a number of neurons and network layers. To make the simulated data closer to the real experimental data, the data set for model training is a time series generated by a computer simulation of the time-correlated single-photon counting process in the presence of Poisson noise. The generated time series is the series corresponding to 20 000 fluorescence lifetime decay histograms uniformly distributed in the range of 100 ns. The data were normalized to eliminate the order-of-magnitude differences, and to avoid large order-of-magnitude differences which would reduce the accuracy of the predictions. The prediction accuracy of the randomly generated 1~100 ns fluorescence lifetime outside the training data set is supposed as the evaluation standard, and the optimal model including 3 layers of LSTM network is selected for the subsequent fluorescence lifetime retrieval. Monte Carlo simulation results indicate that the proposed retrieval algorithm achieves a retrieval accuracy of 95% in the fluorescence lifetime range of 1~90 ns even when the number of photons is 5 000 which is conducive to the fluorescence lifetime imaging. In the case of the same number of photons, the retrieval range is increased by 4.5 times in comparison with the center-of-mass method. Moreover, the proposed method achieves higher retrieval accuracy of the long lifetimes than the traditional least squares method. For the imaging of 32×32 arrays, after several experimental calculations, it is shown that the center-of-mass method can complete the computing in 0.07 s, the least-squares method takes about 77 s, and the proposed algorithm takes about 9.7 s under the conditions of Windows11 (64-bit) operating system, 16 GB memory, and Intel(R) Core(TM) i5-1157G7 processor. The results reveal that neural network not only provides comparable or even better performances but also offers much faster high-throughput data analysis. A shorter time will be used to complete the array imaging when the hardware conditions are improved which provides the possibility for real-time imaging. The proposed algorithm can significantly broaden the fluorescence lifetime reduction range with high retrieval accuracy, thus, it is suitable for accurate fluorescence lifetime retrieval imaging with a single exponential large dynamic range.