Deep Learning–Based Needle Selection and Insertion Depth Determination for Venezia Applicator–Based Cervical Cancer Brachytherapy
Abstract
Purpose
Brachytherapy is a standard component in cervical cancer treatment. Venezia applicators allow addition of limited needle positions to tandem and ovoid procedures, but needle selection and insertion depth largely rely on physician judgement, requiring some level of estimation. This study proposes a deep learning-based approach for automated needle selection and depth determination guided by clinical target volume (CTV) geometry and patient anatomy.
Methods
The model includes (a) an encoder-decoder network for extracting CTV geometric features and (b) a regression network determining insertion depth. The encoder-decoder network adopts a U-Net backbone to encode CTV into geometric features, which are then modulated by an attention mask surrounding a specific needle to emphasize features relevant to the needle. These modulated features are used by the regression network to predict depth. Unused needle positions are defined as zero depths. The model was trained in a supervised manner with clinically recorded depths as the reference. 1,700 needle insertions from 220 Venezia cases were split into 1,324 training and 376 testing samples, each comprising a CTV and the corresponding insertion depth. The model predictions were evaluated against clinical records on a per-needle basis using the absolute difference (AD) from the CTV geometric depth, defined as the distance from the needle-ovoid intersection to the first dwell position.
Results
In 71.81% of cases, the model’s AD from CTV geometric depth was within 3 mm of the clinical-record AD. Among these, 4.26% of all cases had model predictions closer to CTV than clinical records, with needle depths reduced relative to clinical use by 0.52–26.77 mm (median 8.40 mm).
Conclusion
The study demonstrated the feasibility of deep learning for predicting needle selection and depth. Prediction accuracy was limited by inter-physician variability in recorded depths. Future work will use dwell-position-derived depths as reference to better reflect true needle insertion requirements.