Colorectal cancer is a major cause of cancer-related mortality, but early detection and removal of adenomatous polyps can prevent its progression. Optical colonoscopy is the primary screening method, though CT colonography offers a non-invasive alternative with high sensitivity. However, CT colonography cannot reliably differentiate between adenomatous and non-adenomatous polyps, a crucial distinction for therapy management.
Artificial intelligence has emerged as a potential solution, assisting radiologists in more accurately classifying polyps. A study published in European Radiology examines the impact of AI-assisted differentiation on diagnostic accuracy, clinical decision-making and its future role in colorectal cancer screening.
Improving Accuracy and Consistency in Polyp Classification
The integration of AI into CT colonography has been proposed as a means of improving the accuracy of polyp classification, reducing the rate of misclassification and supporting radiologists in their decision-making process. A study evaluating AI-assisted polyp classification assessed radiologists' performance in two phases: an initial reading based solely on current clinical guidelines and a second AI-assisted reading, where radiologists had access to an AI-generated classification of each polyp as adenomatous or non-adenomatous.
The findings demonstrated a notable improvement in diagnostic accuracy when AI was incorporated into the decision-making process. In the unassisted readings, radiologists achieved an accuracy of 76%, a sensitivity of 78% and a specificity of 73% in identifying polyps eligible for endoscopic resection. However, when AI-assisted classification was introduced, accuracy increased to 84%, sensitivity to 85% and specificity to 82%. These improvements indicate that AI can enhance radiologists’ ability to differentiate between polyp types with a higher degree of precision.
Beyond overall accuracy, AI assistance also improved consistency in radiologists' readings. One of the primary concerns in polyp classification is inter-reader variability, where different radiologists may interpret the same imaging data differently. In the study, inter-reader agreement was quantified using Fleiss’ kappa, which increased from 0.69 in unassisted readings to 0.92 in AI-assisted readings. This suggests that AI not only improves individual radiologists' accuracy but also promotes a more standardised approach to polyp classification, reducing discrepancies between different readers.
Impact on Therapy Management and Clinical Decision-Making
The accurate differentiation between adenomatous and non-adenomatous polyps has significant implications for therapy management. Current guidelines recommend the resection of polyps larger than 6mm, with consideration for polypectomy in patients with smaller polyps depending on clinical factors such as age and comorbidities. However, a major challenge in clinical practice is ensuring that polyps requiring resection are accurately identified while avoiding unnecessary procedures for benign polyps.
AI-assisted readings provided radiologists with additional insights, leading to more precise therapy recommendations. By improving diagnostic confidence, AI-supported classification helped to refine decisions regarding which polyps required endoscopic resection. This is particularly relevant for polyps in the 6–9mm size range, where treatment decisions are often more ambiguous. In this category, AI assistance increased classification accuracy from 65% to 77%, sensitivity from 56% to 68% and specificity from 75% to 89%. These improvements suggest that AI can help radiologists more effectively determine which polyps require intervention, potentially optimising patient management strategies.
Furthermore, AI assistance led to changes in radiologists’ decisions in a total of 66 cases. The majority of these changes were beneficial, with 83% of the revised decisions aligning with the histopathological reference standard. This indicates that AI was able to correct initial misclassifications and reinforce accurate decisions, further supporting its role as a valuable second-reader tool in clinical practice.
The Future of AI in Colorectal Cancer Screening
While AI has demonstrated promising results in improving polyp classification at CT colonography, further research is required to establish its full potential in routine clinical practice. One current limitation is that the AI model used in this study relied on manually segmented polyps, requiring radiologists to perform segmentation before AI analysis could be applied. To ensure seamless clinical implementation, future AI systems should incorporate automated polyp segmentation, minimising manual intervention and streamlining workflow efficiency.
Additionally, despite the improvements in classification accuracy, AI-assisted diagnosis is not intended to replace histopathological confirmation. AI serves as a complementary tool that enhances radiologists’ decision-making but does not eliminate the need for histological verification of polyp pathology. Future research should explore the integration of AI into broader colorectal cancer screening programmes, assessing its impact on real-world patient outcomes and evaluating its cost-effectiveness in clinical settings.
The use of AI-assisted differentiation in colorectal polyp classification has demonstrated significant potential in improving radiologists' accuracy, sensitivity and specificity in CT colonography. By reducing inter-reader variability and enhancing diagnostic confidence, AI can refine therapy recommendations, ensuring that patients with adenomatous polyps receive timely intervention while avoiding unnecessary procedures for non-adenomatous polyps. Although further research is necessary to optimise AI models and validate their clinical impact, the findings suggest that AI could become an integral component of colorectal cancer screening in the future.
Source: European Radiology
Image Credit: iStock
