Towards robust deep learning-based autosegmentation in MRI-planned gynecological brachytherapy: Importance of scalable development and comprehensive evaluation
To present comprehensive development and evaluation methodologies for a generalizable deep learning (DL)-driven autocontouring model of standard pelvic organs-at-risk (OARs) in MRI-planned cervical brachytherapy. A curated dataset of 200 3D-MRIs (85% training/validation, 15% testing) including multiple applicator types, varying treated anatomies, and manual contours of OARs (bladder, rectum, sigmoid, small bowel) by 3 physicians was utilized to develop an nnU-Net-based autocontouring model. Iterative tuning was conducted to determine the optimal hyperparameters and enhance evaluation metrics. Model performance was assessed using quantitative metrics, like geometric (e.g., Dice Coefficient (DC) and Hausdorff Distance 95th Percentile (HD95)) and dosimetric (dose-volume histograms (DVHs), dose differences (ΔD2cc)), and then correlated with qualitative physician-review (modified Turing and Likert tests). Geometric metrics were best for bladder (e.g., mean ± SD DC|HD95(mm) 0.93 ± 0.02|2.26 ± 1.07) with greater variability exhibited for small bowel (0.62 ± 0.16|24.90 ± 14.36). Dosimetric comparisons of manual vs predicted contours showed high agreement in DVHs, with mean ΔD2cc <0.60 Gy EQD2 The DL-based autocontouring model, trained on a heterogeneous in-house dataset, demonstrates clinical acceptability for OARs as determined by comprehensive evaluation. It also shows promise for translatability to target contouring, and adaptability to other gynecological (noncervix) brachytherapy applications. Differences in qualitative and quantitative results exist; directionality and magnitude should be considered in clinical usability assessments of brachytherapy autocontouring models.