Congratulations to Prof Ronan Cahill, researchers from the UCD Centre for Precision Surgery, and all those involved in their recent paper, titled, ‘Explainable Endoscopic Artificial Intelligence Method for Real-Time In Situ Significant Rectal Lesion Characterisation – A Prospective Cohort Study’.
Introduction
The accurate assessment of colorectal polyps, particularly those over 2 cm, is crucial for optimizing therapy.(1) Despite advancements, pre-excision evaluation methods remain inadequate.(2) Colonoscopists primarily rely on visual assessment, sometimes aided by spectral-light mucosal illumination, with a diagnostic accuracy of just 40–70%.(3,4) To improve decision-making, tissue biopsies may follow visual evaluation or precede direct excision. However, biopsies, despite offering 80% accuracy, are prone to sampling errors and misclassification.(5) Additionally, they may induce fibrosis, complicating subsequent local excision. Imaging modalities such as MRI and ultrasound fare worse, with approximately 50% accuracy, and are typically performed after colonoscopy. Consequently, many excisions are performed as “excisional biopsies.”
While endoscopic excision can cure some cancers, accurate preoperative oncological assessment could guide biopsy site selection, inform the excision plane (e.g., mucosal, submucosal, or full-thickness), and determine whether radical resection is necessary. Emerging artificial intelligence (AI) tools have demonstrated potential in colorectal lesion identification but face challenges in accuracy and interpretability, particularly for large polyps. Lesions exceeding 2 cm inherently carry at least a 10% cancer risk, even when appearing benign.(6)
Near-infrared (NIR) perfusion-based endoscopic assessment offers a promising alternative, identifying cancers through distinct perfusion patterns associated with beyond-mucosal invasion, angiogenesis, and stromal changes.(7,8) Leveraging advances in computer vision and machine learning (ML), we present an explainable, AI-driven system for real-time cancer discrimination during endoscopy. This system integrates mucosal region of interest (ROI) and full field-of-view (FOV) analyses, providing actionable insights for patients with significant colorectal neoplasia.
Discussion
Our study demonstrates the efficacy of ML-driven dynamic time-fluorescence curve analysis for malignancy discrimination, achieving accuracy comparable to or exceeding traditional imaging and biopsy. A conservative threshold (>0.55) minimized false negatives, facilitating pre-excision planning and patient counselling. The ML-FOV approach, which analyses entire lesions without requiring predefined ROIs, outperformed biopsies and offers enhanced clinical utility and could even aid biopsy targeting, define excision boundaries, and guide surgical strategy.
Errors in perfusion classification were associated with incomplete lesion visualization, prior excisions, and scarring and these findings underscore the need for refinements to address tissue distortion. Interestingly, neoadjuvant therapy did not significantly impact classification accuracy. Expanding the dataset with a higher proportion of benign lesions may improve false-positive rates and enable profiling of non-adenomatous or hyperplastic lesions.
While the study predominantly utilized ex vivo data, clinical validation with trial-grade software is underway. Additionally, the cancer-heavy composition may have inflated malignancy classification accuracy, emphasizing the need for balanced datasets in future studies.
Access the full paper here.
Niall P Hardy MD PhD1, Pol MacAonghusa PhD2, Jeffrey Dalli MD PhD FRCS1, Jonathan P Epperlein PhD2, Paul Huxel PhD3, Mohammad F Khan MD1, Alice Moynihan MD1, Sergiy Zhuk PhD2, Johanna J Joosten PhD4, David Nijssen MD4, Alberto Arezzo PhD5, Juriaan Tuynman PhD6, Peter M Neary MD7, Roel Hompes PhD4, Ronan A Cahill MD1,8