AI-Assisted Endoscopy for Upper Aerodigestive Tract Lesions

Head and Neck Cancer (H&N), Leukoplakia, Squamous Cell Carcinoma, Upper AerodigestiveTract Cancer, Upper Aerodigestive Tract Neoplasms, Upper Aero-digestive Tract (UADT) Neoplasm, Vocal Cord Disease, Vocal Fold Nodules, Vocal Fold Polyp

This is a prospective observational clinical study designed to evaluate the performance of artificial intelligence (AI) algorithms applied to upper aerodigestive tract (UADT) video-endoscopy. The study assesses three main tasks: lesion detection (localization), classification (benign vs malignant), and segmentation of tumor margins. AI algorithms will be applied to endoscopic video data acquired during routine clinical practice without influencing clinical decision-making. The system will process images in real time and store data for subsequent analysis. AI outputs will be compared with physician assessment and reference standard histopathology to evaluate diagnostic performance.

The artificial intelligence algorithms developed will be employed in the analysis of laryngeal lesions for 3 tasks: * Task 1: Computer aided diagnosis (CADx): the algorithm provides a differential diagnosis between benign and malignant neoplasms (binary classification) and the exact histology (multiclass classification). During the UADT video-endoscopy in the outpatient clinic, the physician performs the video-endoscopy and selects and captures n.3 WL and n.3 NBI significant frames of the lesion. The AI model records the classification output of the algorithm that the physician cannot access. The predicted pathologic results will be finally displayed as two different classifications along with the probability of each prediction (0% to 100%) as estimated by the AI algorithm: a first binary classification "neoplastic" or "non-neoplastic," and a second multiclass classification with the exact histology. The physician subsequently, based on the endoscopic examination, will write the suspected diagnosis (benign vs. malignant lesion and the actual histology) in the appropriate patient chart. Next, the physician reviews the screenshot taken and makes sure the lesion is visible in every one of them. Retrospectively, an investigator (blinded to the physician's assessment) will review the AI processed frames with the resulting CADx classifications and mark the AI-processed diagnosis in the patient chart. Once biopsied, the final histology of the lesion analyzed by definitive histopathological examination is recorded in the patient chart by the investigator. The investigators will finally compare the two recorded diagnoses (CADx and physician) with the definitive histology. * Task 2: Computer aided detection (CADe): the algorithm, through the representation of a rectangle (bounding box), localizes the lesion during the video-endoscopy in the outpatient clinic in real-time. During the UADT video-endoscopy, the physician performs the video-endoscopy as for standard-of-care procedure. In parallel, the AI model processes in real-time the endoscopic video and records the output of the algorithm (which the physician cannot access). The physician captures n.3 WL and n.3 NBI significant frames of the lesion. Moreover, n.3 frames where no lesions are visible are captured as negative controls. Later, the physician reviews the screenshot taken and makes sure to label the frames where the lesion is visible as "positive cases" and the frame where the lesion is not visible as "negative cases". The investigators will finally assess if the lesion was detected by the CADe system to define a "true positive". Similarly, to define a true negative, the CADe system should have not output a bounding box in the majority of the "negative cases" frames. * Task 3: Computer aided segmentation (CASe): the algorithm analyzes the neoplasm margins and provides a delineation mask. In the operating room setting, once the lesion to be resected is identified with a 0° telescope, the surgeon captures n.1 WL and n.1 NBI close-up photographs that exemplify the superficial lesion margins. The same procedure is repeated with a 70° optics and other two photographs are acquired. The frames taken are then saved and analyzed by the AI algorithm, which will perform the segmentation task. The surgeon will be blinded to the AI prediction. Later, the surgeon will draw the margins of the lesion according to her/his evaluation of each captured frame. The annotated frame will be saved so that it can be analyzed at a later time. Afterwards, in cases where positive superficial margins are identified by histopathologic examination, the surgeon-designed margins and the AI model ones will be compared to see if there was any difference in the suggested margin.

No related papers found yet.

None listed

Belgium, Italy, Spain