A real-time tumor position prediction based multi-dimensional respiratory motion compensation puncture method

Abstract Objective:
This study proposes a real-time tumor position prediction-based multi-dimensional respiratory motion compensation puncture method to accurately track real-time lung tumors and achieve precise needle puncture.
Approach:
A hybrid model framework integrating prediction and correlation models is developed to enable real-time tumor localization. A Long Short-Term Memory neural network with bidirectional and attention modules (Bi-LSTM-ATT) is employed for predicting external respiratory signals. Subsequently, a backpropagation neural network is constructed to correlate these signals with tumor positions. Tumor trajectory decomposition and the determination of an optimal puncture window based on multiple criteria ensure accurate needle puncture.
Main results:
When the delay time of Bi-LSTM-ATT model is 500 ms, its RMSE, MAE, and R2 are 0.0482 mm, 0.0414 mm, and 97.90% respectively. The correlation model locates lung tumors in 10 cases with a target registration error within 0.74 mm. The proposed puncture method achieves a puncture error ranging from 1.00 mm to 1.32 mm, with an average error of 1.2 mm.
Significance:
The proposed method is validated for its high accuracy and robustness, establishing it as a promising tool for percutaneous biopsy procedures within the lung.