Introduction
High-resolution anoscopy (HRA) is the gold standard for detecting anal squamous cell cancer (ASCC) precursors. It has shown superiority over cytology for detecting high-grade squamous intraepithelial lesions (HSIL). Nevertheless, the visual characterization of these lesions remains challenging. Preliminary studies on applying deep learning models to HRA have revealed promising results. The impact of staining techniques and anal manipulation for the treatment of precursor lesions on the effectiveness of these algorithms has not been evaluated.
Aims & Methods
Our aim was to develop a deep learning system for automatic detection and differentiation of HSIL versus low-grade squamous intraepithelial lesions (LSIL) in HRA images in different subsets of patients (non-stained, acetic acid, lugol, and after manipulation).
A convolutional neural network (CNN) was developed based on 102 HRA exams from 88 patients from a Proctology high-volume center (GH Paris Saint-Joseph, Paris, France). A total of 27,770 images were included, 19,114 images containing HSIL and 8,656 LSIL. A patient-split five-fold cross-validation analysis was performed to assess the network's performance. Partial subanalyses were performed to evaluate the performance of the CNN in the subset of images with no staining (n=2,820), acetic acid (n=13,378), lugol (n=2,195), and after treatment of the anal canal (n=9,377). For these subanalyses, training, and testing datasets comprised 90% and 10% of the total number of images for each class. The sensitivity, specificity, accuracy, positive and negative predictive values, and area under the curve (AUC) were calculated.
Results
The overall accuracy of the CNN in distinguishing HSIL from LSIL during the testing stage was 98.3%. The algorithm had an overall sensitivity and specificity of 97.4% and 99.2%, respectively. The accuracy of the CNN for detecting HSIL vs. LSIL varied between 91.5% (post-manipulation) and 100% (lugol) for the four categories of subanalysis. The AUC ranged between 0.95 and 1.00
Conclusion
The introduction of AI algorithms to HRA may enhance the early diagnosis of ASCC precursors, improving these patients' therapeutic course and prognosis. Staining techniques are essential during HRA exams. Our algorithm showed excellent performance for staining with both Lugol and acetic acid. Moreover, the CNN showed high-level performance in the post-manipulation/treatment images subset. This is extremely important as real-time AI models during HRA exams can help guide local treatment or detect relapsing disease.
References
1. Islami F, Ferlay J, Lortet-Tieulent J, Bray F, Jemal A (2017) Inter-
national trends in anal cancer incidence rates. Int J Epidemiol
46(3):924–938. https://doi.org/10.1093/ije/dyw276
2. Smittenaar CR, Petersen KA, Stewart K, Moitt N (2016) Cancer
incidence and mortality projections in the UK until 2035. Br J
Cancer 115(9):1147–1155. https://doi.org/10.1038/bjc.2016.304
3. Ebisch RMF, Rutten DWE, IntHout J et al (2017) Long-lasting
increased risk of human papillomavirus-related carcinomas and
premalignancies after cervical intraepithelial neoplasia grade 3:
a population-based cohort study. J Clin Oncol 35(22):2542–2550.
https://doi.org/10.1200/jco.2016.71.4543
4. Deshmukh AA, Suk R, Shiels MS et al (2020) Recent trends in
squamous cell carcinoma of the anus incidence and mortality in
the United States, 2001–2015. J Natl Cancer Inst 112(8):829–838.
https://doi.org/10.1093/jnci/djz219
5. Maugin F, Lesage AC, Hoyeau N et al (2020) Early detection of
anal high-grade squamous intraepithelial lesion: do we have an
impact on progression to invasive anal carcinoma? J Low Genit
Tract Dis 24(1):82–86.