# Developing an automatic decision‐assistance tool to choose proton/photon radiotherapy for patients with prostate cancer

## TL;DR

This study creates an automatic tool to help choose between proton and photon radiotherapy for prostate cancer patients based on predicted tissue damage.

## Contribution

A novel automatic decision method using deep learning to select optimal radiotherapy techniques for prostate cancer.

## Key findings

- The system achieved 93.75% accuracy in predicting the correct treatment modality for 48 patients.
- No significant differences were found between manual and model-predicted NTCP values for the rectum wall.
- The decision method showed high AUC values (0.88), indicating strong predictive performance.

## Abstract

It is important to guide staff in choosing appropriately between photon and proton radiotherapy. This study develops an automatic decision method to select the most clinically beneficial radiotherapy technique (proton or photon) for patients with prostate cancer. An automatic decision method was developed to help staff in choosing appropriately between photon and proton radiotherapy for patients with prostate cancer.

Forty‐eight patients with prostate cancer were enrolled. First, photon and proton dose prediction models (M
ph and M
pr) were trained using reference plans from previous patients’ data. Second, the predicted values of V
6300cGy (rectum wall) were obtained using the trained models, M
ph and M
pr, and these values were used to calculate the Normal Tissue Complication Probability (NTCP). Finally, if the photon NTCP exceeded 10%, the proton NTCP was calculated, and the difference (ΔNTCP) between the two was used to guide treatment selection. The accuracy of the decision support system was evaluated by comparing dose distributions, NTCPs, and decision outcomes between manual and automatic plans using paired t‐tests.

The deep learning (DL) model showed a mean absolute error (MAE) of 4.60 ± 1.80 for the rectum wall in the photon group and 3.64 ± 1.27 in the proton group. There was no statistically significant difference in V
6300cGy (rectum wall) between manual plans and model predictions (photon group p = 0.594, proton group p = 0.057). Similarly, no significant differences were observed in NTCP values for the rectum wall (photon group p = 0.383, proton group p = 0.100). The system correctly predicted the treatment modality in 45 of 48 cases, resulting in an accuracy rate of 93.75%, with AUC values for the decision method at 0.88.

The proposed automatic decision method matches dose distributions, accurately calculates NTCPs, and supports precise radiotherapy technique selection, enhancing the clinical efficiency.