Artificial intelligence classification of rectal neoplasia by endoscopic fluorescence perfusion analysis

Disordered vascularity is a hallmark of carcinogenesis. Fluorescence microperfusion heterogeneity may discriminate malignant transformation within significant (> 20 mm) rectal polyps enabling in-situ endoscopic classification via machine learning (ML) methods to inform clinical care. Patients referred for transanal management of rectal neoplasia were recruited. Indocyanine green was administered intravenously and near-infrared (NIR) endoscopic video recorded. Videos were processed with bespoke fluorescence quantification software producing intensity-timeseries plots for neoplastic and normal regions of interest in the same patients. Plot features were extracted to train/test ML classification algorithms, including coefficient of variation (CV), reporting cancer characterisation sensitivity, specificity and accuracy. 190 video recordings from 182 patients (57.9% with cancer) from six cancer centres provided usable dataset (91% of 201 consenting patients). Overall, the software accurately tracked and detailed NIR perfusion features from, on average (SD), 74.7% (25.3) of annotated regions of interest over the five-minute recording phase. The sensitivity/specificity/accuracy rates of traditional endoscopic biopsy (n = 172), MRI (n = 139) and expert surgeon opinion (n = 190) at surgery were 70.8%/100%/81.7%, 85.4%/ 44.1%/72.7% and 79.1%/80%/79.5% respectively. In comparison, trained ML sensitivity/specificity/ accuracy was 77.6%/39.8%/61.1% and 73.5%/48.2%/62.6% with base and CV featured algorithms respectively. Combining point of care clinical data (specifically MRI and clinicians’ preoperative predictions) with the ML algorithms improved sensitivity/specificity/accuracy to 86.0%/71.1%/79.5% and 82.2%/74.7%/79.0% respectively. Malignant transformation precipitates discriminant perfusion patterns, in a manner exploitable digitally, that indicate cancer presence in significant rectal polyps. Combining clinical indicators appears to improve classification accuracy further, especially specificity. Trial registration: Future of Colorectal Cancer Surgery (FOOCCuS1). Clinicatrials.gov. NCT04220242. Clinicaltrials.gov/study/NCT04220242. CLASSICA: Validating AI in Classifying Cancer in Real-Time Surgery. Clinicaltrials.gov. NCT05793554. Clinicaltrials.gov/study/NCT05793554.