Random Forests for Automatic Paranasal Sinus and Nasal Cavity Detection in CT Images

Sinusitis and other pathologies of the nasal cavity and paranasal sinuses are common diseases. Physicians need segmentations of the structures for diagnosis, operation planning, learning models and more. The current practice is manual or sometimes semi-automatic segmentation of the computed tomography (CT) scans. This time-consuming, tedious task is not practical for clinical routine. An automatic segmentation of the paranasal sinuses and nasal cavity in CT images is therefore desirable, but extremely hard to accomplish because of the high variance and complexity of the structures. An often applied approach for difficult segmentation tasks is to use bounding boxes (BBs) of the structures as regions of interest (ROIs) to limit the search space for the segmentation to a small area in the image. This thesis therefore proposes a method for an automatic detection of the paranasal sinuses and nasal cavity in CT images that predicts axis-aligned bounding boxs (AABBs) of the structures. The BBs can then be used as ROIs in a possible subsequent segmentation. The relative locations of the nasal cavity and paranasal sinuses are strongly regularized by the human anatomy. It therefore makes sense to use this knowledge for the localization of the structures of interest. Thus, a random forest (RF) is proposed, which performs a simultaneous regression of the BBs of all structures of interest. Thereby it implicitly learns the relative locations of the structures directly from the training data. Intensity based features that capture spatial relations are applied. For the detection of the paranasal sinuses the forest is trained on 71 and tested on 29 CT scans. A subset of 30 images of the dataset is furthermore used for the simultaneous detection of the paranasal sinuses, nasal conchae and nasal septum. The dataset contains images with many different resolutions and sizes, often pathologies are present. With a mean Intersection over Union (IoU) of 0.27 for detection of the paranasal sinuses and 0.21 for all structures at once, the results are still improvable. Though, in some cases good detection results are achieved. Furthermore, the results show that the proposed method is able to learn the relative positions of the nasal cavity and paranasal sinuses from the training data. This offers interesting potential for the detection of the nasal cavity and paranasal sinuses.