Hybrid contour and distance-based loss for enhanced bladder and prostate segmentation in CT images

Segmentation of bladder and prostate in Computed Tomography (CT) scans is an important task in clinical practice for aiding in diagnosis, monitoring conditions, and identifying cancerous lesions, which affect around 2.1 million patients suffering from lower urinary tract symptoms. Traditionally, radiologists manually segment these structures on CT scans to identify and locate unusual tissues, shapes, and worrying areas. Despite its importance, it has been considered a backbreaking task due to: i) difficulty, ii) time-consuming, and iii) subject to inter- and intra-observer variability. This research proposes a novel approach for delineating the bladder and prostate in CT scans. For that, a new hybrid loss function was implemented to enhance bladder and prostate segmentation, by gathering the information of the segmentation, contour, and distance map. Moreover, this study also includes a comparative analysis of DL architectures to assess the improvement of the segmentation detail. The proposed method showed strong performance on the AMOS22 and PAEC datasets, containing 292 training, 82 validation, and 43 testing images. For the AMOS22 dataset, mean average surface distance (aSSD) for the bladder and prostate were 1.5 ±1.54mm and 1.65 ±1.78mm, respectively, and the 95th percentile of Hausdorff Distance (HD95) values were 7.31 ±11.34mm and 5.04 ± 3.42mm, respectively, corresponding to mean Dice scores of 88.76 ± 9.02% and 85.90 ± 12.93%. The method on the PAEC dataset yielded mean aSSDs of 1.73 ± 1.35mm and 3.26 ±1.67mm with reductions of 66.47% and 36.08% for the bladder and prostate, respectively, with average Dice scores of 90.80 ± 7.91% and 80.57 ± 7.07%, and average HD95 values of 7.82 ± 7.80mm and 8.67 ± 3.18mm. The obtained results suggest the potential clinical interest of this AI model to assist doctors in performing screenings of the bladder and prostate.