Fig. 3.

Boxplots comparing the mDTA for the four model configurations trained using the best performing loss function, ExpLogLoss, and the deviation between doctors for reference (blue boxes). For this figure, lower values indicate better segmentations. Configurations using 3-channel input (3 R&T) outperform the single-channel counterparts (1 R&T) in all soft tissue OARs. Models with traditional transpose convolutions (T) produce marginally better segmentations, with the best-performing model highlighted in green.