Speaker
Description
We present a compact 3D neural network for segmentation of maxillary sinuses from CT scans in children with chronic rhinosinusitis. It comprises 12 convolutional layers totaling 50k parameters, minimal compared to most 3D medical-imaging architectures, and completes the prediction in a few seconds on an average CPU, making the tool accessible for most clinicians. The sinus volume and its fraction occupied by inflammatory changes are estimated with an RMSE of 0.5 cm$^3$ and 1.5 %pts, respectively, which allows for precise longitudinal monitoring. The network includes no dropout, no pooling, and no padding, the latter preventing layer-wise injection of meaningless zeros. We utilize a custom normalization that does not average the data but employs the same running statistics in both prediction and training, making the model insensitive to overall contrast and local artifacts. The network is fully convolutional and translation-invariant. It has an inner receptive field of 18x18x18 voxels to detect sinus walls and an outer receptive field of 88x88x88 voxels to provide a broader context. The internal data flow is designed to minimize the number of mappings that the network must learn. It operates in a reduced resolution of 1x1x1 mm$^3$, which still allows for high precision thanks to the use of fuzzy labels accounting for the partial-volume effect. The model was trained and tested on a dataset of 92 scans collected and manually annotated specifically for this study.
