Comparison of LSTM and RNN model performance in predicting F-18 NaF kinetics in prostate cancer bone metastasis based on PET/CT
DOI:
https://doi.org/10.61511/crsusf.v2i2.1951Keywords:
long short term memory (LSTM), NaF, prostate, recurrent neural network (RNN)Abstract
Background: PET/CT imaging using F-18 NaF is an important modality for evaluating bone metastasis in prostate cancer. The accuracy of this tracer kinetics prediction can improve monitoring of therapeutic response. Although RNN and LSTM have been used for modeling sequential data, no comprehensive study has specifically compared their performance for predicting F-18 NaF uptake in prostate cancer, despite the clinical importance of cost and PET/CT service availability. Methods: This study analyzed data from nine patients in the NaF PROSTATE dataset (TCIA). SUVmean was extracted from PET/CT imaging series. Bidirectional LSTM and RNN models with dropout layers were developed. Evaluation was performed using R², RMSE, MAE metrics, a nd kinetic analysis via biexponential curve fitting to assess the biological plausibility of predictions. Findings: Evaluation of model LSTM demonstrated superior performance than RNN. Kinetic curve analysis confirmed that LSTM was able to reproduce uptake and clearance patterns more stably and physiologically than RNN, which showed fluctuations. These findings are consistent with the theoretical advantage of LSTM in handling long - term dependencies. Conclusion: LSTM is proven to be superior to RNN in predicting the kinetics of F-18 NaF in prostate cancer bone metastases, both statistically and clinically. Its accuracy and stability support its potential application in molecular imaging and therapy monitoring. Novelty/Originality of this article: This study provides quantitative evidence of LSTM's superiority over RNN for predicting F-18 NaF kinetics, using an innovative validation approach through kinetic curve analysis that enriches clinical assessment beyond conventional statistical metrics.
Published
Issue
Section
Citation Check
License
Copyright (c) 2025 Sani Inuwa Sani, Mohammad Sidiq Cahyana, Haryanto

This work is licensed under a Creative Commons Attribution 4.0 International License.