In this study, the characteristics of the MLC position error under non-heterogeneous conditions were explained using only the dosiomics indices because they were statistically significant among the gamma, SSIM, and dosiomics indices. The determined dosiomics indices were used for the predictive MLC position error model. The clinical relationship of significant indices and prediction MLC position error model was examined using DVH.
For gamma index results, even in the case of MLC position systematic 1.0 mm shift, it calculated 0.905 for the C-shape easy case and 0.710 for the C-shape hard case. These results show that the dose distribution was affected by plan complexity, and the gamma index is considered to be low. However, there is no significant difference within 3% in the DVH index for PTV and cord. Then, the gamma index has a low correlation with the DVH parameter. The SSIM index tendency according to the random MLC displacement is relatively small and irregular compared with the SSIM index according to the systematic MLC displacement (Table 3, Fig. 4, and Supplemental Table 3). For the systematic error, only the luminance SSIM subcomponent was more sensitive than the other SSIM subcomponents for the MLC position error. However, its sensitivity was also relatively small compared with the dosiomics indices (Supplemental Table 4). For the random error, the gamma and SSIM indices did not convey any trend of the MLC position error.
Among the gamma, SSIM, and dosiomics indices, statistically significant indices representing the characteristics of MLC position error were extracted from dosiomics. In Class-I, GLCM_Energy and GLCM_Entropy_log10 were selected as common significant indices for all predictive models. Both indices belong to the GLCM, representing the dose distribution with co-occurring pixel values at one offset. In the subtracted dose map between the error-free and error data, GLCM_Energy represents the uniformity of gray-level voxel pairs, and GLCM_Entropy_log10 represents the randomness of gray-level voxel pairs. They had a strong Spearman’s rank correlation (> 0.8). Therefore, it was shown that uniformity decreased while at the same time randomness increased with co-occurring pixel values at one offset. In Class-II, GLRLM_LRHGE was selected as common significant indices for all cases, and it belongs to the GLRLM, representing the size of the homogenous run. GLRLM_LRHGE shows the distributions with long homogenous runs with high gray levels. As a result, the feature of long homogenous runs with high gray levels were presented for the MLC systematic position error. In Class-III, GLCM_Energy was chosen as the common significant dosiomics index as in Class-I except for C-shape hard cases where the significant index was gray-level zone length matrix (GLZLM) gray level_nonuniformity (GLNU). Comparing this result with previously published results, our result that more than half of the statistically significant indices belonged to the GLCMs was consistent with that of Chaoqiong Ma et al. 2021 [2]. In addition, GLRLM_LRHGE, a common significant index that can detect systematic errors, and GLCM_Energy or GLCM_Entropy_log10, a common significant index that can detect random errors, are consistent with the results of the paper published by Landon S. Wootton et al., 2018 [21]. These two studies were performed using different devices and techniques than ours, and while our study did not include clinical variations, these two studies did include them. Nevertheless, the significant dosiomics indices we found were consistent with those in these studies. These consistent results confirmed that the significant indices we found were a basic index that characterizes the MLC position error regardless of the measurement device, technique, and clinical variation.
The predictive models for MLC position errors were developed using only dosiomics indices in Class-I, Class-II, and Class-III. The gamma and SSIM indices were disregarded because they were not dedicated adequate weights for developing the predictive models, compared to the dosiomics index. The final error prediction models using the significant dosiomics indices describing the characteristics of the MLC position error exhibited excellent performance with AUC > 0.9, and accuracy, sensitivity, and specificity ≥ 0.8 (p < 0.05), except for the C-shape hard case of Class-III. The more the plan complexity, the more distributed was the influence of the MLC position error; therefore, it was estimated that the accuracy, precision, and specificity values other than AUC and sensitivity values were less than 0.8.
As a result of the DVH analysis, it was found that the relative percentage difference in systematic error increased almost linearly. This means that the larger the systematic error in which the MLC position is offset in one direction, the larger the area where the dose distribution differs. For a random error in which the MLC position is offset in a random direction, the DVH analysis result showed that the relative percentage error increased with different trends depending on the structure’s location. The reason was that the structures in different locations were affected differently for random errors because MLC position errors occur in a random direction. As a result, the systematic error affected the size of the homogenous run in the dose distribution difference. By contrast, the random error affected the voxel pairs in the dose distribution difference. These results confirmed that the common significance index indicating the characteristics of the differential dose distribution for systematic errors was GLRLM_LRHGE, and that of random errors was related to GLCM_Energy. Regarding the complexity of the plan, the DVH result showed that the more complex the plan, the greater the relative percentage difference. However, no common significant index representing the dose distribution difference was found. This is probably due to the greater scatter of nonuniformity showing a significant common index as the complexity of the plan increases. These characteristics make different dose distributions depending on the plan complexity, and the significant dosiomics index also varies due to the resulting random dose distribution differences. This phenomenon was confirmed in the examples of the C-shape easy case and C-shape hard case. Each of them had two different significant indices, GLCM_Energy and GLZM_GLNU. While GLCM_Energy selected from the C-shape easy case considers a pixel pair with a specific value, the significant index GLZM_GLNU chosen from the C-shape hard case considers the connected voxels. As a result, depending on the complexity of the plan, the DVH provides information on the dose-volume histogram but cannot show the localized texture differences between dose distributions. In the case of dosiomics analysis, it can provide essential additional information that DVH cannot offer because it can show differences in local texture of dose distribution depending on the complexity of the plan.
The results of our study are summarized in three key findings as follows. First, the discovery of principle common significance indices (GLCM_Energy, GLRLM_LRHGE) represents the characteristics of systematic and random MLC position errors and can be used as reference indeices for future clinical applications. Second, the prediction MLC position error models developed using statistically significant dosiomics indices showed superior performance (AUC > 0.9). Third, it was confirmed that the results of DVH in the primary state were related to dosiomics analysis, which represented the characteristics of the dose distribution difference due to the MLC position error. In addition, it has been demonstrated that in cases with different plan complexity, dosiomics analysis can give more critical information that can be added to the information in the DVH. The results of our study can ultimately provide information related to the characteristics of localized texture using dosiomics analysis. This information could be made more clinically useful by providing additional information to the gamma index and DVH commonly provided by devices and technologies such as log file-based QA or electronic portable imaging devices (EPID) for dosimetric impact assessment.
Although our results demonstrated the novel clinical findings mentioned above, our study has two limitations. First, the dosiomics index was extracted from the subtracted dose distribution, but the gamma and SSIM indices were extracted from dose distributions of the error-free and simulated error instead of their subtracted dose distribution. Second, since only one type of error (MLC position error) was studied under the condition that the factors affecting the error analysis are removed, the direct application of the result to various clinical process is limited. In future, we plan to investigate detection sensitivity utilizing dosiomics, SSIM, and DVH according to diverse error scenarios and complexities in heterogeneous environments.