Table 1.
The overall characteristics of the patient dataset. Distribution of class labels of each dataset was displayed. Left/right sinuses are counted separately.
Fig 1.
Three-dimensional U-Net (N = 4 down-sampling levels) for maxillary sinus segmentation.
The encoder comprises four levels; each level applies two 3 × 3 × 3 convolutions with ReLU activation and batch normalization, followed by 3D max-pooling. Feature channels by encoder level are 32, 64, 128, and 256. The decoder mirrors the encoder with transposed convolutions for up-sampling and skip connections to the corresponding encoder features, followed by two 3 × 3 × 3 convolutions (ReLU and batch normalization). A final 1 × 1 × 1 convolution with softmax produces voxel-wise probability maps of the maxillary sinus. Training optimized a Dice-based loss with standard spatial and intensity augmentation.
Fig 2.
A series of axial slices from PNS CT scans demonstrating maxillary sinus anatomy and calcification patterns.
a) normal sinus or chronic sinusitis, representing clear sinuses or sinus opacifications without abnormal calcifications, b) dense peripheral calcification (dystrophic), denoting denser calcification areas often associated with non-fungal sinusitis, and c) central punctate calcification pattern seen in fungal sinusitis.
Fig 3.
Predicted results of the maxillary sinus segmentation model.
The model accurately delineated the area of interest, successfully segmenting challenging regions with ambiguous boundaries due to topological changes.
Fig 4.
Predicted results of the fungal calcifications detection model.
Regions suspected of calcifications were identified and measured using bounding boxes. While the model accurately localized the areas of interest, there was a slight tendency for higher false positive cases, such as implantation artifacts and small bone structures.
Table 2.
Performance summary of the segmentation, detection, and classification models resulted during training process of each model. The table shows the metrics for segmentation (Dice loss, DSC), detection (CIoU loss, BCE loss), and classification (training/validation accuracy and loss), providing an overview of the models’ performance.
Table 3.
Performance metrics of the calcification pattern classification model across internal validation set, temporal external test set, and geographic external test set. Overall accuracy and balanced accuracy (macro recall) are reported to provide a comprehensive evaluation of the model’s classification performance across fungal and non-fungal sinusitis cases.
Fig 5.
Receiver Operating Characteristic (ROC) curves for multi-class classification across three experiments.
a) internal validation, b) temporal external test set, and c) geographic external test set. High discrimination performance was observed for Class 2 and Class 3, while the discrimination for Class 1 was relatively low.
Fig 6.
Confusion matrices for a) internal validation, b) temporal external test set, and c) geographic external test set.