Expanding the use of artificial intelligence in non-invasive colorectal cancer screening

Abstract

There is a growing interest in and utilization of artificial intelligence (AI) to improve polyp detection during colonoscopy. While colonoscopy is the gold standard for colorectal cancer screening, there remain significant barriers to achieving population screening goals. Alternative non-invasive screening methods remain important adjuncts, and the use of AI is expanding. A review of the available literature regarding the use of AI in improving colorectal cancer screening using non-invasive methods was performed, including PubMed, MEDLINE, and Cochrane Databases. Researchers are using AI to improve non-invasive colorectal cancer screening tests, including blood-based tests, radiographs, computed tomographic colonography, and capsule endoscopy. Most of these tests, particularly the blood-based and X-ray tests, still lack the necessary sensitivity to be true alternatives to colonoscopy. Breathonomics/metabolomics is evolving as a screening tool with the use of AI. Capsule endoscopy and computed tomographic colonography, already strong alternatives to colonoscopy, are impressively enhanced with the use of AI. AI strengthens non-invasive colorectal cancer screening modalities and shows promise in broadening alternative screening options. The current literature features studies with small sample sizes and retrospective data, limiting the ability to support its reliable application in clinical practice.

Key Words: Artificial intelligence; Colorectal; Cancer; Screening; Non-invasive

Core Tip: The use of artificial intelligence (AI) has made a significant impact in the development and/or enhancement of non-invasive colorectal cancer screening methods. AI is improving blood-based and radiographic testing and progressing novel methods like breathonomics. AI-aided colon capsule endoscopy is particularly adept at identifying polyps and precancerous lesions. Further prospective studies remain necessary to clarify its utility in broadening its application in clinical practice.

INTRODUCTION

Colonoscopy is the gold standard screening tool for colorectal cancer, but it is not accessible to all patients. While extremely safe and well-tolerated, colonoscopy is an invasive procedure that carries some risk. Medical comorbidities and anatomic constraints can preclude some patients from safe colonoscopic screening. High cost and resource scarcity are limiting factors for other patients[1,2]. Even among those who undergo initial screening with colonoscopy, there remain high non-adherence rates for recommended repeat surveillance or screening intervals. To this end, there remains a need for reliable alternative screening tools. A variety of non-invasive colorectal cancer screening modalities are currently in use, ranging from fecal and-blood based tests to more novel testing, such as breathonomics.

In an effort to improve screening, there has been a growing demand for the incorporation of artificial intelligence (AI) in colorectal screening. These systems are meant to optimize tasks and projects that would otherwise be performed by medical staff, with problem-solving abilities that mimic the human nervous system[3,4]. Lesion detection software is already commonly used during colonoscopy, improving endoscopists’ adenoma detection rates. Lesser known are the efforts to strengthen the non-invasive colorectal screening tools currently available to patients by employing machine-based and deep learning systems.

Fecal testing is the most widely adopted alternative to colonoscopy, but the use of AI in this area is limited. Stool-based screening tests also have poor sensitivity for advanced adenomas and carry significant false positive rates[5,6]. In this review, we explore the use of AI for the other non-invasive colorectal cancer screening methods. As the incidence of colon cancer increases, providers should be knowledgeable of the promising advances being made in non-invasive screening, particularly radiographic modalities such as computed tomography (CT) colonography. We describe the available non-invasive and minimally invasive tests and the current literature centered around AI ‘s role in improving its performance. An electronic search across PubMed, MEDLINE, and Cochrane databases was performed to identify peer-reviewed studies published between 2005 and 2025. Searches were performed using keywords such as “artificial intelligence”, “colorectal cancer screening”, and “non-invasive colorectal cancer screening”. Relevant articles were selected from among 2305 PubMed responses, 100 MEDLINE responses, and 312 Cochrane Database responses. We included available English language studies that featured any form of virtual AI in the setting of the development or application of any non-invasive or minimally invasive colorectal cancer screening test. Studies that did not report findings of a test’s accuracy or validity were not included. Articles exclusively studying colonoscopy or fecal testing were not included.

RESULTS

Screening modalities

Blood tests: Blood-based testing for colorectal cancer screening relies on subtle changes in blood work that can signal increased risk for or the presence of colorectal cancer when applied to large datasets of patients with confirmed colorectal cancer. These tests may serve as useful risk-stratifying tools, particularly for patients who are hesitant to undergo colonoscopy or where colonoscopy is less accessible. Two validated blood-based risk assessment/screening tests exist currently in the literature. Developed in Israel, the ColonFlag model is a machine-learning model using age, sex, and complete blood count values over a defined period to calculate a predictive risk of asymptomatic colorectal cancer[7]. The predictive ability of the model was tested using a large cancer database, which allowed for analysis of pre-diagnosis parameters of patients with confirmed colorectal cancer. The model was then tested in two independent data sets. A meta-analysis of ten studies that evaluated the ColonFlag test reported a wide variability in sensitivity (3.91%-35.4%). The test produces an area under the curve (AUC) range of 0.736 to 0.82 (a score of 1 represents perfect test discrimination, whereas a score of 0.5 indicates poor diagnostic capabilities, no better than random guessing). The test resulted in a specificity of 82.73% to 94%, positive predictive value of 2.6% to 9.1%, and negative predictive value of 97.6% to 99.9%. It is particularly adept at predicting more advanced disease as compared to adenomas. The model was noted to heavily rely on age as a predictive measure[8]. This has been externally validated numerous times - not all of these studies are included individually in this review.

The University of Oxford developed its own similar blood-based predictive model (BLODDTRACC) using an advanced “supercomputer” meant to evaluate the two-year risk of colorectal malignancy. This study showed similar results to the ColonFlag test. They reported c-statistic scores by male and female groups, and compared these results to ColonFlag. Scores were reported as BLOODTRACC = 0.751, ColonFlag = 0.762 for males and BLOODTRACC = 0.763, ColonFlag = 0.761 for females[9]. Tokutake et al[10] describe a machine-learning model based on a single-center dataset in Japan that could predict the presence of colorectal polyps. Similarly, a machine-learning model developed at the University of Michigan demonstrated that analysis of demographic information and lab values could successfully predict GI tract cancer at 3-years (including colorectal cancer) with an AUC of 0.750[11].

Breathonomics

Cancer cells have distinct metabolisms that produce unique metabolite profiles. Some of these metabolites are volatile organic compounds that are present in exhaled breath. This is the basis of a novel screening tool called breath volatolomics or breathonomics. A breathing device is transfigured to collect and analyze data about the composition of volatile organic compounds and other metabolites in a patient’s breath, most often using a process called gas chromatography-mass spectrometry (although there are others). Machine learning is currently being used in many phases of breathonomics work, including data acquisition, pre-processing, and feature extraction. There are ongoing efforts to broaden the use of AI in this field with deep learning and neural networks for more clinical applications[12]. One particular diagnostic model of exhaled volatile organic compounds reported an AUC of 0.89 (95% confidence interval: 0.76-0.91), accuracy of 0.78, specificity of 0.77, and sensitivity of 0.86 for detecting colorectal cancer[13]. These results are promising but limited, with a lack of large validation studies and concerns regarding standardization.

Urine metabolomics

Similar to breathonomics, urine metabolomics analyzes the composition of urine metabolites. A study out of the University of Alberta created a machine-learning predictive model based on the urine compositions of 988 patients who underwent a colonoscopy. Using the urine metabolite profiles and the colonoscopy results, the model was able to predict the presence of an adenoma with a sensitivity of 64% and specificity of 65%[14]. This modality is still in its infancy regarding its use in colorectal cancer screening, lacking large-scale validation studies, but it has already shown great promise in early disease detection for other pathologies.

X-Ray

A study out of Taiwan evaluated the use of Kidney-Ureter-Bladder (KUB) studies in predicting the presence of colorectal cancer[15]. A deep learning system was used to analyze the anatomic details of the colon on KUBs, which allowed for the identification of subtle colon wall thickening, dilation, and signs of obstruction. The model was tested in an internal validation set (patients from two hospitals in Taiwan) and an external validation set of a similar patient cohort. This group reported an AUC of 0.738, sensitivity of 61.3%, and specificity of 74.4% for the internal validation set. The external validation set data results were more tempered, with a reported AUC of 0.656, sensitivity of 47.7%, and specificity of 72.9% for the external validation set, reflecting the challenges in its implementation as a screening tool in the general population. The predictability was best for high-grade colon cancers, reporting an AUC of 0.744[15].

CT colonoscopy

CT colonography or “virtual colonoscopy” is an alternative screening tool, often used for patients who are unable to undergo or have incomplete colonoscopies. It involves a bowel prep to clear the colon and administration of air or CO₂ gas in the colon prior to CT imaging. This modality has been transformed with the use of computer-aided detection software and other AI systems. Computer-aided software assists in detecting subtle mucosal changes and flat lesions, some of the easiest to miss on colonoscopy and CT imaging. We identified five studies that use various forms of AI to aid in polyp detection and differentiation (Table 1)[16-20]. Two studies specifically focused on improving the detection of polyps[20,21]. Another two were designed to improve recognition of adenomatous vs non-adenomatous polyps[16,18]. AI-aided software was used to detect early carcinomas (T1 Lesions), although this study had a very small sample size[17]. It is consistent across the relevant studies that AI-assisted CT colonography is more accurate and sensitive in detecting polyps than standard imaging (Table 1)[16-20].

Table 1 Artificial intelligence and computed tomography colonography review.

Ref.	CT colonography
Song et al[16], 2014	Virtual pathological model exploring texture features to differentiate neoplastic from non-neoplastic lesions	148 colon lesions; AUC: 0.74 (using the image intensity alone) to 0.85 (considering the gradient and curvature images)
Taylor et al[17], 2008	Computer-aided detection software to detect flat early colon carcinoma	24 flat T1 tumors; CAD detected 20 (83.3%), 17 (70.8%), and 13 (54.1%) of the 24 cancers at filter settings of 0, 0.75, and 1
Grosu et al[18], 2025	AI-assisted differentiation of adenomatous and non-adenomatous colorectal polyps as compared to standard radiology reading	59 patients, 118 polyps; AI-assisted readings with higher accuracy (76% ± 1% vs 84% ± 1%), sensitivity (78% ± 6% vs 85% ± 1%), and specificity (73% ± 8% vs 82% ± 2%) in selecting polyps eligible for polypectomy (P < 0.001)
Alkabbany et al[19], 2022	AI-based fusion of 2D projections with 3D colon representation images to generate new synthetic images used to train a RetinaNet model to detect polyps	49 patients, 59 Polyps; 94% f1-score and 97% sensitivity
Endo et al[20], 2025	Deep-learning-based AI algorithm developed to improve the detection of polyps	92 lesions in the interval validation dataset; sensitivity of 0.815, 0.738, and 0.883 for lesions ≥ 6 mm, 6 mm to 10 mm, and ≥ 10 mm, respectively

AUC: Area under curve; CAD: Computer-aided detection; AI: Artificial intelligence; 2D: 2-dimensional; 3D: 3-dimensional.

Colon capsule endoscopy

Colon capsule endoscopy is the only colonoscopy alternative that offers direct visualization of the colonic mucosa via a camera-equipped pill that is swallowed. Technically considered minimally invasive, it is performed without the need for anesthesia but still requires bowel preparation. Limitations include the time necessary and specialized training needed to evaluate the images. The introduction of AI offers the potential to overcome some of these challenges and improve data processing. Without AI, colon capsule endoscopy impressively detects polyps and malignant tumors with comparable accuracy to colonoscopy itself. Sensitivity for polyps > 6 mm ranges from 79% to 96% and is higher for larger polyps. The rate of detection of colorectal cancer is approximately 93%[21]. Deep learning appears to further improve its ability to recognize precancerous and cancerous lesions. We identified seven studies with models designed to detect findings ranging from intraluminal blood and ulcers to tumors. These studies boast sensitivity and specificities in the high 90s (Table 2)[22-28].

Table 2 Artificial intelligence and colon capsule endoscopy review.

Colon capsule endoscopy	Deep-learning models using convolutional neural networks to better detect colonic abnormalities
Ref.	Model designed to detect
Ribeiro et al[22], 2025	Ulcers and erosions	124 CCE exams; AUC: 1.00; accuracy: 99.6%, sensitivity: 96.9%, specificity: 99.9%; overall accuracy: 99.6%
Mascarenhas et al[23], 2022	Intraluminal blood and mucosal lesions	124 CCE exams; mean sensitivity: 96.3% and specificity: 98.2%; mucosal lesions - sensitivity: 92.0%, specificity: 98.5%; blood - sensitivity: 97.2%, specificity: 99.9%
Mascarenhas et al[24], 2022	Protruding lesions	124 CCE exams; AUC: 0.99; accuracy: 95.3%, sensitivity: 90.0%, specificity: 99.1%, PPV: 98.6%, NPV: 93.2%
Saraiva et al[25], 2021	Protruding lesions	24 CCE exams; AUC: 0.97; sensitivity: 90.7%, specificity: 92.6%, PPV: 79.2%, NPV: 96.9%
Blanes-Videl et al[26], 2019	Polyps, compared to trained endoscopists	250 patients; accuracy: 96.4%, sensitivity: 97.1%, specificity: 93.3%
Gilabert et al[27], 2022	Polyps, and evaluate the reviewing time compared to the classical linear review software	18 studies; reviewing time was reduced by a factor of 6, and polyp detection sensitivity was increased from 81.08% to 87.80%
Yamada et al[28], 2021	Polyps and cancers	15933 CCE images; AUC: 0.902; accuracy: 83.9%, sensitivity: 79.0%, specificity: 87.0%

CCE: Colon capsule endoscopy; AUC: Area under curve; PPV: Positive predictive value; NPV: Negative predictive value.

DISCUSSION

Colonoscopy is the ideal colorectal screening method. It provides a magnified direct view of the colonic mucosa to visualize potential lesions and provides the ability for immediate removal of precancerous lesions. Although very safe and well-tolerated, it is invasive and requires bowel preparation and often sedation. Colonoscopy is not universally available, and the need for alternative screening methods is undeniable. The goal is to balance invasiveness, which allows for more accurate screening, and cost-effectiveness that improves accessibility. The ideal screening test would be easy to perform, without much procedural risk, and with high sensitivity and specificity in detecting advanced premalignant lesions and early-stage cancers.

Non-invasive or minimally invasive screening methods are generally hindered by low sensitivity/specificity for advanced precancerous lesions. The promise of AI is in its potential to enhance the accuracy of non-invasive tests, thus broadening legitimate screening methods available to patients. AI-based adenoma detection software for colonoscopy is already widely in practice and enhancing the performance of endoscopists in significant ways. Deep neural networks and machine learning algorithms are being utilized in biomedical problem-solving and have begun to permeate colorectal cancer screening database studies[29-31].

Blood and stool-based tests are the least invasive types of colorectal cancer screening tools. Stool tests offer acceptable specificity rates for colorectal cancer but are plagued by low sensitivity in detecting advanced precancerous lesions, with sensitivities under 50%. This is true even for multi-target stool DNA tests, such as Cologuard[32,33]. The ColonFlag blood test, which utilizes machine-learning, demonstrated sensitivities and positive predictive value similar to fecal occult blood tests in retrospective studies, but has high specificity and negative predictive value for the presence of colorectal cancer[8]. Some patients may be less averse to blood-based tests, making it potentially more accessible. Blood tests likely have the most potential to serve as a risk-assessment tool and an adjunct to stool-based tests[33].

While AI-led advances in breathonomics and metabolonics are exciting and have true potential for early colorectal cancer detection, there remain concerns around their clinical utility. Metabolite biomarkers are often low-concentration in exhaled breath, and can be distorted by environmental factors (smoke, air pollution). There is also a lack of standardization in the collection process and volume analyzed within the available studies. The future of its clinical application remains unclear. Novel approaches in AI-enabled X-ray imaging analysis yielded exciting results in the detection of colorectal cancer using KUBs. This is a cost-effective approach to radiographic screening, but it appears to be most effective for left-sided and advanced cancers. AI-assisted software improves the accuracy of CT colonography and can enhance its ability to identify small and flat lesions, but this also results in increased radiation burden for patients. Colon capsule endoscopy is expected to present the highest sensitivity and specificity of all the minimally-invasive screening methods, and deep learning models consistently enhance its already impressive ability to detect colonic lesions. However, patients don’t avoid a bowel prep, and the cost-effectiveness as compared to colonoscopy has not been established[34].

With the primary goal of increasing participation in colorectal cancer screening, we embrace AI and the ways it improves non-invasive colorectal screening methods. Ethical considerations surrounding the use of AI in healthcare continue to be at the forefront of the conversation. These systems are designed to support medical staff and the patients they care for by boosting the quality of the work performed. We all share the responsibility of ensuring that AI’s increasing presence in healthcare is guided by the principles of safety, accountability, and equity. While the findings discussed in this mini-review are promising, additional larger-scale prospective studies are necessary before promoting the broad implementation of AI for the above studies in the clinical setting.

CONCLUSION

Many of the available studies included in this review are limited by small sample sizes and their retrospective nature. This is understandable given the need for large datasets in validation studies of AI-based algorithms and systems. Currently, AI improves non-invasive testing but does not yet overcome its major limitations in significant ways. These tests ultimately may serve an important role in helping stratify patients by risk, in order to better guide providers in areas with limited resources for gold-standard colonoscopic screening.