Author: Majd Antaki, Rohini Bhatia, Gayoung Kim, Yosef Landman, Junghoon Lee, Akila N. Viswanathan 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Physics and Advanced Development Elekta 🌍
Purpose: Brachytherapy is a standard radiation therapy approach for cervical cancer, which directly delivers radiation source to the tumor using catheters. Treatment planning requires identification of the target tumor and catheters from CT and/or MRI to ensure therapeutic precision, which is typically done manually, thus labor-intensive. While deep-learning models have been explored to automate these tasks, their performance depends on the quality of training data and lacks mechanisms to assess uncertainty in prediction. Therefore, we propose an uncertainty-aware deep-learning model generating prediction and its uncertainty.
Methods: The uncertainty-aware dual convolution-transformer UNet (U-DCT-UNet) comprises a combined convolution-transformer encoder, a transposed-convolution decoder, and an uncertainty block. The encoder and decoder predict voxel-wise probabilities for segmentation, while the uncertainty block estimates confidence level. The segmentation module firstly localizes the region of interest (ROI) of high-risk clinical target volume (HR-CTV) using naïve DCT-UNet without the uncertainty block. Then, U-DCT-UNet segments detailed HR-CTV boundary from the ROI volume and calculates the prediction uncertainty. The U-DCT-UNet is also applied to predict catheters. From the model segmentation, individual catheters are distinguished using 3D connectivity and spatial structure, then further optimized using uncertainty values to eliminate false positives.
Results: The segmentation module was trained on 223 MRIs, and tested on 27 cases with four sets of manual contours obtained from three different radiation oncologists. The module achieved dice similarity coefficient of 0.71±0.11 and 95th percentile Hausdorff distance of 9.31±4.47 mm compared to the consensus manual segmentations with the estimated uncertainty well-reflecting the multi-observer variability. The catheter detection module was trained and tested on 28 and 7 MRIs, and successfully detected 98.5% of catheters with shaft and tip errors of 0.74±0.32 mm and 2.52±2.04 mm, respectively.
Conclusion: The experimental results demonstrate that uncertainty estimation provides valuable insights into model predictions, enhances performance, and supports clinicians in making informed decisions.