ESTRO meets Asia 2024 - Abstract Book

S68

Interdisciplinary – Brachytherapy

ESTRO meets Asia 2024

Brachytherapy (BT) is vital to the cure of locally advanced cervix cancer, but workflows are increasingly complex with image-guided approaches. Automating BT processes may alleviate burden, streamline treatments, and improve outcomes. Prediction models developed using artificial intelligence (AI) are becoming common. In this practical report of our team’s journey in the discovery and development of deep-learning (DL)-based autosegmentation of organs-at-risk (OAR) for cervix BT using real-world MRI imagesets, we review common process considerations and pitfalls.

Material/Methods:

A narrative review was undertaken of our team’s lived experience in the development of a DL-based autosegmentation model for cervix BT, with focus on common OAR structures (bladder, BLD; rectum, REC; sigmoid, SIG; small bowel, SB). Practicalities were assessed by phase: pre-training, training, validation, and application.

Results:

We identified nearly 300 cases of image-based curative gynecological BT boost featuring MRI that were treated (2016 – 2024) after the latest clinical practice change in OAR contouring. Exclusions were applied: those planned on CT-only, those who did not proceed to BT delivery, those without the 4 OAR structure sets of interest, and non cervix cancer diagnoses. For patients treated with multiple fractions (PDR or HDR), each MRI imageset was considered separately. In total we retained 200 T2-weighted 3D MRI scans from time of intracavitary +/- interstitial (IS) BT, which all featured an intrauterine tandem along with a range of vaginal applicator components (ring, Utrecht ovoids, Venezia lunar ovoids, Geneva ovoids, or multichannel cylinder; Elekta, Sweden). Ground truth (GT) was taken to be the OAR structures contoured by 3 experienced radiation oncologists (RO), in nearly equal proportions (35:34:31), on 1 mm slices. All cases were pooled to develop an autosegmentation model using DL (using a modified nnUNet) after assessing 2 other AI techniques in-depth: 70% for training, 15% for validation, 15% for testing. Issues were encountered at each phase, including pre-training. Examples include: under- or over contouring by ROs (including partial OAR contours, not necessarily systematic nor RO-specific), motion artifacts on GT scans, and significant differences in GT OAR volumes (inter- or intra-patient variability). During training, more granular understanding of scan quality and contour quality emerged, for instance overlapping OAR structures (infrequent) required manual adjustments to GT, slice by slice. Hyperparameters for DL modelling had to be worked out. Early results also highlighted other challenges: with also under- or over-contouring by the model, or imagined OAR contours. Iterative learning was adopted to improve the model. Qualitative and quantitative measures of model performance were monitored. Application learnings are ongoing, with periodic evaluation by ROs and other BT stakeholders, for clinical-level of quality.

Conclusion:

Clinical teams can develop DL-based autosegmentation algorithms aligned with local BT practices. Some curation is inevitably required for the training MRI set, but learning from a broad range of treated anatomies and BT applicator scenarios is achievable. Clinical practice-specific DL-based segmentation, to mimic real-world contouring, may be within reach.

Keywords: cervix brachytherapy, deep learning autocontouring

466

Digital Poster