Deployment and assessment of a deep learning model for real-time detection of anal precancer with high frame rate high-resolution microendoscopy

Abstract

Anal cancer incidence is significantly higher in people living with HIV as HIV increases the oncogenic potential of human papillomavirus. The incidence of anal cancer in the United States has recently increased, with diagnosis and treatment hampered by high loss-to-follow-up rates. Novel methods for the automated, real-time diagnosis of AIN 2+ could enable "see and treat" strategies, reducing loss-to-follow-up rates. A previous retrospective study demonstrated that the accuracy of a high-resolution microendoscope (HRME) coupled with a deep learning model was comparable to expert clinical impression for diagnosis of AIN 2+ (sensitivity 0.92 [P = 0.68] and specificity 0.60 [P = 0.48]). However, motion artifacts and noise led to many images failing quality control (17%). Here, we present a high frame rate HRME (HF-HRME) with improved image quality, deployed in the clinic alongside a deep learning model and evaluated prospectively for detection of AIN 2+ in real-time. The HF-HRME reduced the fraction of images failing quality control to 4.6% by employing a high frame rate camera that enhances contrast and limits motion artifacts. The HF-HRME outperformed the previous HRME (P < 0.001) and clinical impression (P < 0.0001) in the detection of histopathologically confirmed AIN 2+ with a sensitivity of 0.91 and specificity of 0.87.

Optical imaging modalities for the detection of head and neck cancer margins

“…55 The versatility and compact design of HRME have facilitated its deployment across a spectrum of anatomical sites, spanning from HNC to esophageal squamous cell neoplasia, as well as HPV-associated cervical and anal cancers. [54][55][56][57][58] In 2012, Vila et al demonstrated the feasibility of HRME in the discrimination of benign and malignant mucosa in exvivo HNSCC specimens from 38 patients, achieving a Sn 98.3% and Sp 90.9% with high interobserver reliability. 55 Miles et al then applied HRME to the visual intraoperative margin in 33 HNSCC patients, displaying HRME-generated images to blinded observers.…”