Residual Dense Blocks for Extreme Foreground Imbalance in Brachytherapy Applicator Segmentation

Brachytherapy is a vital medical intervention for treating gynaecological malignancies in females. During brachytherapy, the precise detection of a radioactive probe’s location through imaging is crucial to safeguard critical organs. Unfortunately, datasets containing such information are often limited. Moreover, the probe’s location is confined to a small region, causing a significant imbalance in foreground-background ratios for segmentation tasks. Consequently, training deep learning models capable of real-time probe segmentation presents a challenging yet essential requirement for medical applications. In this paper, we address this issue by annotating CT scan samples from 140 patients undergoing brachytherapy where the probe locations are annotated by expert medical physicists. Along with the dataset, we also propose a novel architecture that incorporates residuals in residual dense blocks within an encoder-decoder framework for near-real-time probe segmentation from CT scan images. Our research demonstrates that the proposed architecture effectively handles the highly unbalanced segmentation problem and outperforms existing solutions, offering promising advancements in the field of brachytherapy for gynecological malignancies.