Search

The paranasal sinuses are a group of four paired air-filled spaces that exist superior and laterally to the nasal cavity. The sinuses are considered to provide an immunological role along with mucus production and air humidification. These sinus tissues are susceptible to obstructive malformation, inflammatory, infectious, and neoplastic processes, sometimes requiring surgical management. Surgical management has moved from open approaches to minimally invasive endoscopic sinus surgery (ESS). This surgical approach, although effective, has a reported complication rate of up to 5%, with cerebrospinal fluid leak due to perforation of the skull base reported to occur in up to 1% of cases. The regions of the skull base at risk of perforation include the cribriform plate medially and the fovea ethmoidalis laterally, with the lateral lamella of the cribriform plate (LLCP) the thinnest structure and hence most susceptible to iatrogenic injury. Thus, examination of pre-operative radiographic imaging, which has transitioned from plain X-rays to the currently preferred computed tomography (CT) modality, is critical for patient safety when performing ESS as at-risk anatomy can be readily identified and protected. However, in radiologist reports of CT scans, landmarks of surgical importance are often not commented upon, requiring thorough and careful re-review by the operating surgeon. These deficiencies not only place the patient at greater risk as concerning structures may be missed but also means clinically relevant information is not reported for detailed attention during surgical review. Recent studies using computer vision based artificial intelligence (AI) tools show significant promise in their ability, after appropriate training, to identify and classify anatomy of interest and to streamline workflows for both radiologists and surgeons in screening for significant anatomic irregularities to highlight for clinician review. Through experimental research, a computer vision based convolutional neural network (CNN) AI model was developed to measure the height of the LLCP and then to categorise it as per the clinically relevant Keros classification system. The research included a systematic review of the published literature to analyse the use of CNNs in analysing radiographic imaging in rhinology, identify algorithms showing promise and their current applicability to clinical practice (Chapter 3). The review indicated a scarcity of published papers and of the 7 papers that used CT scans, all of them used varying CNN models and transfer learning techniques to classify different rhinologic anatomy. Notably, studies show that the CNN algorithms exhibit weaker performance when tested on external image datasets suggesting limitations in generalising trained algorithms for use in different institutions and populations. Therefore, future studies that test the performance and applicability of these AI tools should include validation steps, with external datasets. While these CNN algorithms were able to identify pathologic structures or variations of normal anatomy, they did not measure the size of these pathologies, requiring the need to develop and train a measuring algorithm in this study. The multi-layered CNN AI algorithm developed in this study and trained using randomly selected coronal CT hemi-slices from the prepared image-set was then used to measure the LLCP height in the remaining hemi-slices within the image-set (test images). This CNN AI measured a total of 114 out of 140 LLCPs correctly to within 1mm of the ground truth, an accuracy of 81.4% with an overall mean absolute error of 0.61mm (95% CI = 0.52-0.70mm). The mean absolute percentage error (MAPE) was 15.8% and the Pearson correlation coefficient was 0.58 (95% CI: 0.46 – 0.68). This demonstrates moderately strong agreement between the ground truth, measurements completed by two surgical residents, and the CNN AI measurements. When the measured LLCPs heights were assigned a Keros stage, the algorithm obtained an accuracy of 73.6%, area under the curve of 0.79 (95% CI: 0.71 – 0.86) and a staging agreement kappa statistic of 0.44 (95% CI: 0.29 – 0.59) between the ground truth and the trained CNN, representing moderate correlation (Chapter 4). While various studies have shown that CNN AI algorithms with appropriate training can identify and classify anatomy of interest, the findings of this research have made an original contribution towards using CNN AI algorithms to measure clinically important structures, such as the LLCP, on preoperative CT scans, which can then be to classified according to established staging systems to aid in operative planning. The use of machine learning (ML) tools in the assessment of pre-operative paranasal sinus CT scans and their potential clinical integration warrants further investigation.