Patients
The study protocol was approved by the Human Ethics Review Committee of the Jikei University School of Medicine, Tokyo, Japan (approval number: 32-036 [10111]), which waived the requirement for informed consent owing to the retrospective nature of the study.
We retrospectively evaluated and enrolled 53 patients (male, n = 33; female, n = 30; mean age, 51.2 ± 12.6 years) who underwent endoscopy sinus surgery at our hospital from 2018 to 2021, including 21 patients diagnosed with IP by pathological examination and 32 patients with chronic rhinosinusitis with nasal polyps (CRSwNP). Video images were used to show the nearly bloodless condition prior to manipulation; forceps were not included in the endoscopic image.
Endoscopy videos
All videos were taken using a rigid 4.0-mm nasal endoscope of 0° angles and a camera head (Olympus Medical System Corp., Japan, and Karl Storz Endoskope, Germany). The main frame rate of the video was 119.88 frames per second, and the resolution was 1920 × 1080 pixels.
Neural network
We adopted the MobileNet-V2 network, a relatively compact network comprising 88 layers with a fixed input image size of 224 × 244 pixels and 3,538,984 learning parameters.
Training
The original images were augmented to 6 million images. Augmentation was performed randomly without considering the balance between the number of original images of each patient. During training, the DNN models learned using images resized to 224 × 224 pixels. In each epoch (training cycle), 120,000 images were randomly selected from the aforementioned 6 million images, and a total of 50 epochs were repeatedly performed to train one DNN model. This 50-epoch training procedure was performed with eight datasets, and eight models were generated using one learning set (learning set:evaluation set ratio, 7:1). As DNN models exhibit differences in ability each time they are trained using a large amount of data generated via augmentation from a small number of patients, we created 25 training sets to verify the fluctuations in accuracy of each model. Consequently, 200 models were generated (8 datasets × 25 = 200 models).
Evaluation
We used square images resized to 224 × 224 pixels. The eight models obtained in each learning set were used as a single evaluation set, and predictions for the 25 evaluation sets were performed as both single-image-unit-based and patient-unit-based predictions. Single-image-unit-based prediction was performed on each single image, whereas patient-unit-based prediction was performed in two ways–continuity analysis and five second (5-s) scoring analysis–with image arrays sequentially aligned according to the order on the original video stream for each patient. In addition to single-model predictions, 25 sets of ensemble predictions combining 24 of the 25 models were used to evaluate the accuracy of the aforementioned predictions. Image-unit-based prediction was simply performed by image-by-image predictions for individual images, whereas patient-unit-based prediction was performed using a continuity analysis and 5-s score analysis for all image sets extracted from a single video of each patient.
Continuity analysis
A continuity analysis was one of our original methods to predict if patients were positive or negative for IP. This method initially evaluates the individual images extracted from the video stream individually and subsequently judges whether a patient is positive or negative for IP based on the number of consecutive positive images in the original video stream.
Five-second scoring analysis
The 5-s scoring analysis was also an original method for the aforementioned purpose. This method judges whether a patient is positive or negative based on the maximum sum of scores obtained from consecutive images in a 5-s video stream.
Diagnostic examination by otolaryngologists
All 53 cases were visually diagnosed by 25 otolaryngologists at our hospital. The target videos were the exact same videos used by the AI for training evaluation, unedited at full length. The otolaryngologists included the surgeon; therefore, the primary surgeon for all eligible cases was anonymized. Hence, surgeons were unable to identify the cases they operated on. The percentage of correct diagnoses was compared with that obtained by the AI. As a secondary item, the correct diagnosis rate for otolaryngology was also examined separately by years of practice experience. The skills of the aforementioned 25 otorhinolaryngologists were classified as follows: entry, < 5 years; intermediate, 4–10 years; and veteran, > 10 years.
