Abstract:
Objective To automatically extract and quantitatively evaluate the dosimetric indicators of nasopharyngeal carcinoma intensity-modulated radiotherapy plans to provide reference for clinical and scientific research practice with Python language and DICOM RT files.
Methods A retrospective analysis was conducted on the clinical data of 21 patients (13 males and 8 females, aged (48.7±21.6) years) with nasopharyngeal carcinoma who underwent radiotherapy in the First Affiliated Hospital of Nanjing Medical University from January 2022 to October 2022. The maximum dose (Dmax); minimum dose (Dmin); mean dose (Dmean), minimum dose covering 98%, 50%, and 2% of the volume (D98%, D50%, and D2%); percentage of volume covered by 95%, 100%, 105%, and 110% of the prescribed dose to its total volume (V95%, V100%, V105%, and V110%) of each target; minimum dose covering 1 cm3 and 1% of the volume (D1cc, D1%); Dmax, Dmin and percentage of volume covered by 30 Gy dose to the total volume (V30 Gy) of the organs at risk were counted in the Monaco 5.11 system by manual methods. The conformity index (CI), homogeneity index (HI), and gradient index (GI) were calculated using corresponding formulas. RT Structure and RT Dose files were exported. The same automatic extraction and quantitative evaluation of radiotherapy plans were completed using Python language (Python methods), and the findings were compared with the quantitative results from manual methods. Data from the two methods were compared by paired samples t-test (homoscedasticity).
Results Compared with those obtained by manual methods, the differences in dosimetric indicators such as Dmax, Dmin, Dmean, D98%, D50%, D2%, V95%, V100%, V105%, V110%, and HI of each target region obtained by Python methods were all less than 3%. The Dmin difference in PTV1 was the largest at (2.21±0.17)%, followed by the Dmin difference in PTV2 at (1.75±0.13)% and the V105% difference in PGTVnd at (1.67±0.09)% , and the differences were not statistically significant (t=0.093, 0.071, 0.065; all P>0.05). The maximum differences in CI and GI were 0.007±0.108 and 0.016±3.958, respectively, and the differences between these two values were not statistically significant (t=0.011, −0.034; both P>0.05). Nevertheless, Python methods could calculate the S_index and HI3. The evalution results of organs at risk show that the Dmax of optic nervus and chiasma calculated by Python methods was slightly lower than that computed by manual methods, and the differences were greater than 1 Gy, and the differences were statistically significant (t=−2.909, −3.420, both P<0.05). The results of the remaining quantitative evaluations for the organs at risk obtained by the two methods were nearly the same, and the differences were less than 0.5 Gy or 0.5%. In terms of evaluation time for radiotherapy plans, Python methods only took 1–3 min for a single case, which was greatly shorter than the 30–50 min required by manual methods.
Conclusions The dosimetric data automatically extracted by the Python methods is basically consistent with those collected by manual methods, but the evaluation time is greatly reduced, the evaluation indicators are more comprehensive, and the evaluation methods are more widely used, which can provide technical reference for the automatic evaluation of radiotherapy plans based on artificial intelligence.