A qualitative, quantitative and dosimetric evaluation of a machine learning-based automatic segmentation method in treatment planning for gastric cancer

Abstract

Purpose

To investigate the performance of a machine learning-based segmentation method for treatment planning of gastric cancer.

Materials and methods

Eighteen patients planned to be irradiated for gastric cancer were studied. The target and the surrounding organs-at-risk (OARs) were manually delineated on CT scans. A machine learning algorithm was used for automatically segmenting the lungs, kidneys, liver, spleen and spinal cord. Two radiation oncologists evaluated these contours and performed the required editing. The accuracy of the auto-segmented contours relative to manual outlines was evaluated by calculating the dice similarity coefficient (DSC), Jaccard score (JS), sensitivity and precision. VMAT plans were initially created on manual contours (MCPlans) and, then, on edited and unedited auto-segmented contours (AC_edPlans). Dose parameters of the OARs and target volume derived from the different treatment plans were statistically compared.

Results

The 24.6 % of the auto-segmented contours were acceptable and 40.5 % needed changes related to stylistic deviations. Minor editing was applied in 34.1 % of these contours. The mean values of the DSC, JS, sensitivity and precision associated with the comparison of the manual outlines and the contour set including edited and unedited auto-segmented contours were 0.91–0.97, 0.84–0.94, 0.92–0.97 and 0.91–0.97, respectively. No significant differences were found for fifteen out of eighteen examined dosimetric parameters derived from MCPlans and AC_edPlans (p > 0.05). These parameters from the MCPlans agreed well with those from AC_edPlans based on the Bland-Altman test.

Conclusions

The qualitative, quantitative and dosimetric analysis highlighted the clinical acceptability of a machine learning-based segmentation method for radiotherapy of gastric cancer.