# Quantifying the impact of tumor size and motion on 4DCT‐4DCBCT image registration accuracy using machine learning and statistical analysis

## TL;DR

This study uses machine learning and statistical analysis to determine how tumor size and motion affect the accuracy of 4DCT-4DCBCT image registration in radiotherapy.

## Contribution

The study introduces a systematic method to quantify the impact of tumor size and motion on registration accuracy using Random Forest and ANOVA.

## Key findings

- Tumor size and SI motion were the most significant factors affecting registration accuracy.
- Registration accuracy deteriorated significantly with small tumors and high SI motion.
- Respiratory cycle and AP/LR motions had no statistically significant impact.

## Abstract

This study systematically quantifies the effects of five variables—respiratory cycle, tumor size, and motion amplitudes in the superior‐inferior (SI), anterior‐posterior (AP), and left‐right (LR) directions—on the registration accuracy between four‐dimensional computed tomography (4D CT) and four‐dimensional cone‐beam CT (4D CBCT) images, thereby providing a theoretical basis for optimizing registration strategies in image‐guided radiotherapy (IGRT).

A CIRS 008A dynamic phantom fitted with 1 and 3 cm tumor inserts was utilized to simulate various respiratory motion scenarios by manipulating respiratory cycles (T = 0, 2, 4, and 8 s) and three‐dimensional motion amplitudes (SI, AP, and LR ranging from 0 to 15 mm, with AP and LR limited to 0, 1, and 5 mm). Corresponding four‐dimensional images were acquired using a GE Discovery RT CT simulator and a Varian VitalBeam linear accelerator. Rigid registration between the 4D CT and 4D CBCT images was subsequently performed using the Varian imaging system, with registration quality quantitatively assessed via the Dice similarity coefficient (DSC). Furthermore, a Random Forest regression model was employed to determine the relative importance of each factor, and multifactor analysis of variance (ANOVA) was conducted to verify statistical significance.

The Random Forest analysis indicated that, for the overall registration average intensity projection, the factors were ranked in order of importance as follows: tumor size (0.509), SI motion (0.315), respiratory cycle (0.094), LR motion (0.055), and AP motion (0.028). In the maximum intensity projection, tumor size (0.722) was found to have a particularly significant impact. The multifactor ANOVA further supported these findings, demonstrating that tumor size (p < 0.001) and SI motion (p < 0.001) have a highly significant influence on registration quality, whereas the respiratory cycle and AP/LR motions did not reach statistical significance (p > 0.05). Notably, when the tumor size was small (1 cm) and accompanied by considerable SI motion (>10 mm), registration accuracy markedly deteriorated, with the greatest variability observed under these conditions.

This study demonstrated that the registration quality between 4D CT and 4D CBCT images was significantly influenced by both tumor size and the amplitude of motion in the SI direction.