Preprocessing Design Space Exploration for Panoramic Dental Segmentation

Accurate segmentation of dental panoramic radiographs is crucial for computer-aided diagnosis; however, the field lacks a systematic understanding of how preprocessing choices interact. Unlike prior work that varies a single factor, we present a comprehensive study of the joint effects of data augmentation, edge detection, and loss functions on performance and efficiency. Using 120 CBCT panoramas with six anatomical classes, we assess 64 configurations across seven U-Nets. Top preprocessing pairs markedly outperform the baseline. Sobel with Optical Distortion and CE–Dice improves the composite validation score by + 9.46 points over (NoAug + NoEdge + CE). Laplacian with Optical Distortion and dice–focal loss adds + 9.44 points. We validate generality on a 1000-image Tufts dataset [15]: baseline achieves Acc \(95.13\%\) , Dice \(98.25\%\) , Val Acc \(94.04\%\) , Val Dice \(89.21\%\) , mIoU \(96.66\%\) (Val mIoU \(81.54\%\) ). The best pipeline, Optical Distortion with dice–focal–loss, improves Val Dice by + 1.38% and Val mIoU by + 1.88% over baseline. An accuracy–efficiency analysis using Dice/Wh identifies CE + Dice with Sobel and Optical Distortion as the strongest performance–compute balance among tested options. Rather than a single global optimum, the results surface top-performing, model-agnostic configurations and an energy-normalized metric that clarifies how combined preprocessing choices improve dental panoramic segmentation.