Could artificial intelligence-powered colonoscopies change the future of colorectal cancer screening?

The preferred method of CRC screening is FOBT. Most studies indicate that the traditional guaiac-based FOBT is inferior to the FIT, making FIT the preferred method for CRC screening. More recently, the introduction of the ColonView FIT test has further advanced screening accuracy and convenience[ 45 - 54 ]. To improve adenoma detection methods, various studies have assessed different AI-based enhancements, such as logistic regression (LR), support vector machine (SVM), neural network (NN), RF and GBM[ 55 ]. All tested AI models demonstrated superiority in colorectal neoplasia screening, primarily evaluated through the adenoma detection rate (ADR), as adenomas are the precursor of CRC. ADR remains the key performance metric and a critical endpoint for reducing CRC incidence[ 56 ]. Among the various models assessed, GBM/RF/NN algorithms were identified as the most effective[ 55 ]. The diagnostic accuracy of various AI models surpassed that of conventional low-risk/high-risk FOBT and FIT tests. Specifically, GBM, RF, NN, SVM, and LR models achieved sensitivities of 85%/82%/68%/60%/55% specificities of 74%/72%/76%/69%/77%, and overall efficiencies of 79%/76%/73%/65%/67%, respectively[ 55 ]. In CRC screening, the preferred initial method is a fecal-based test, with patients who test positive subsequently undergoing a full colonoscopy. In the United States, however, screening often begins directly with colonoscopy. In both approaches, a major problem is inadequate bowel preparation, which occurs in approximately 25% of cases and frequently necessitates a repeat procedure. Additional limitations include longer procedure times and a reduced ADR, reported to be as low as 48% under suboptimal conditions[ 57 - 59 ]. The potential role of AI in improving bowel preparation for colonoscopy remains under investigation. Recent studies have evaluated ChatGPT-generated pre-colonoscopy preparation prompts, which can be tailored to individual patients and healthcare providers[ 60 - 62 ]. As anticipated, AI has also been applied to develop bowel preparation scoring systems[ 61 ]. Moreover, analyses of prompt disagreement have identified several common issues, including: Overly long or short preparation times, dietary misunderstandings, improper timing or medication use, unclear pre-procedure clearance instructions, confusing sequence of steps, mismatched preparation types, excessive medical jargon, lack of readability at a sixth-grade reading level, and unclear guidance on fasting requirements[ 62 ]. When various specialists, including senior physicians, fellows, junior doctors and representatives of the American Gastroenterological Association, evaluated the ChatGPT-generated pre-colonoscopy bowel preparation prompts, they found the material to be easy to understand, scientifically accurate and effective in addressing common patient questions about colonoscopy. Notably, post-procedure outcomes demonstrated significant improvement, with surveillance management compliance reaching nearly 90% and accuracy approximately 85%. Furthermore, several studies have shown that providing ChatGPT with contextualized medical guidelines enables it to generate highly accurate recommendations for colonoscopy screening intervals[ 63 - 67 ]. For instance, two commonly used bowel preparation instructions given to patients were found to exceed the recommended sixth-grade reading level, scoring 8.5 and 7.8 on the Flesch-Kincaid reading scale and 10.5 and 8.6 on the Gunning Fog Index, respectively. In comparison, the AI-generated preparation prompts were rated as scientifically accurate, more patient-friendly, and better aligned with clinical standards. However, they were noted to be slightly too complex, indicating room for further refinement[ 62 ]. The ADR is a critical quality indicator in colonoscopy, as higher ADRs are directly associated with reduced overall risk and burden of CRC. Improving ADR relies heavily on maintaining high-quality colonoscopy performance. However, despite ongoing advances, approximately 27% of polyps are still missed during colonoscopy, largely due to non-endoscopic factors, such as inadequate bowel preparation and varying levels of endoscopist expertise[ 68 ]. Several studies have explored strategies to address this issue, with recent research focusing on the integration of AI into colonoscopy[ 69 - 78 ]. One such method is ALPHAON, an AI-based CADe algorithm with exceptional performance in polyp detection. It has demonstrated an accuracy of 0.97 [95% confidence interval (CI): 0.96-0.99], sensitivity of 0.91 (95%CI: 0.85-0.97), specificity of 0.99 (95%CI: 0.97-0.99), and an area under the ROC curve of 0.967. Additionally, it enhances outcomes across experience levels. Among trainees, it achieved an accuracy of 0.95 (95%CI: 0.93-0.96) and sensitivity of 0.80 (95%CI: 0.74-0.86), while among experts, performance was slightly higher, with an accuracy of 0.96 (95%CI: 0.94-0.97) and sensitivity of 0.84 (95%CI: 0.79-0.90). Although ALPHAON demonstrates impressive diagnostic performance, it still has notable limitations. Currently, it cannot classify polyps or provide pathological information, though future developments may allow its transition into a CADx system. Another limitation involves false-positive detections, such as bubbles, mucosal folds, and tags, which can increase withdrawal time, lead to unnecessary biopsies, and contribute to operator distraction and fatigue[ 79 - 81 ]. Multiple randomized controlled trials have demonstrated that incorporating CADe significantly improves both ADR and polyp detection rates compared to conventional colonoscopy alone[ 82 - 85 ]. However, while numerous trials have evaluated different AI algorithms at varying levels, the findings remain variable and inconclusive, underscoring the need for further validation and standardization(Table 1 )[ 86 - 108 ]. Trials using computer-aided detection system for colorectal lesions[ 86 - 108 ] AQCS: Automatic quality control system; GI: Gastrointestinal; CAD: Computer-aided diagnosis; ENDO-AID: Endoscopy computer-aided diagnosis system; AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate; AMR: Adenoma miss rate. Some studies are advancing FocuU2Net, an innovative bi-level nested U-structure integrated with a dual-attention mechanism[ 109 ]. The model integrates focus gate modules for spatial and channel-wise attention and residual U-blocks with multi-scale receptive fields for capturing diverse contextual information. Comprehensive evaluations on five benchmark datasets: Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and EndoScene demonstrate Dice score improvements of 3.14% to 43.59% over state-of-the-art models, with an 85% success rate in cross-dataset validations, significantly surpassing prior competing models with sub-5% success rates (Table 2 )[ 110 ]. Compared to more recent methods such as SR-AttNet, PolypSegNet, CSAP-UNet, I2-UNet, segment anything model 2 (SAM2) and segment anything in medical images (MedSAM), FocusU2Net maintains competitive performance, highlighting its relevance and effectiveness in the evolving research landscape[ 110 ]. When discussing deep learning-based approaches for polyp segmentation approaches, several architectures are noteworthy[ 110 ]: Traditional encoder-decoder architectures (U-Net, U-Net++, ResU-Net++, ResUNet, MultiResUNet) offer key advantages including efficient feature propagation through skip connection, volumetric skip connection with deep supervision, nested skip connections with residual learning, enhanced gradient flow, and multiscale feature fusion. However, these architectures can face certain challenges, particularly in capturing broader contextual information, preserving fine structural details, and accurately segmenting complex or irregularly shaped polyps. Additionally, multi-scale feature handling remains a technical limitation. Performance comparison of studied models[ 110 ] The best values. The second-best values. Dice score: A metric used to measure the resemblance of two groups; IoU: A metric used to evaluate the accuracy of annotation, segmentation and object detection algorithms; Rec: Recall the true positive rate; Prec: Precision the degree of accuracy. Modified encoder-decoder architectures (DilatedRes-fully convolutional network, PolypSegNet, SR-AttNet, U2 Net, I2 Net, PSPNet, DC-UNet, MSRFNet, BA-Net) benefit from features such as dilated residual convolution (a deep-learning tool to look at a wider area of an image without losing details) to improve fully convolutional network, skip connections via multi-scale context fusion, stretch-relax attention with feature-to-mask extraction, nested U-structure to capture rich contextual information, dual-path U-Net for low-level details, high-level semantics, pyramid scene parsing network for pixel-level prediction, efficient residual CNN for improved gradients, multi-scale feature fusion by multiple dual shot face detectors blocks, boundary-aware modules to explicitly find edge-details, but fail to excel because of poor generalization, risk of overfitting (learning the training data too well, but performing poorly on new, unseen data), feature redundancy, semantic gap in encoding–decoding. Transfer learning-based approaches (DoubleUNet, HarDNet-MSEG, Inception U-Net) represent another important category of deep learning models for polyp segmentation. These methods leverage pretrained feature extractors to enhance performance and efficiency, offering several advantages: Atrous spatial pyramid pooling for improved multi-scale context capture in semantic segmentation networks, stacked U-Net architectures, high-performance backbones with cascaded decoders for refined feature extraction, and multi-modality fusion with deep encoding, enabling integration of information from multiple data sources. However, these models also present notable limitations, including limited interpretability, high computational cost, overfitting and adaptability issues. Transformer-based approaches (SwinE-Net, TransUnet, SwinPA-Net) have some advantages, such as EfficientNet with Swin transformer [vision transformer (ViT) architecture] for global context, transformer-CNN hybrid encoder with U-Net Decoder, Swin transformer with local pyramid attention module. Limitations include lack of inductive biases (guiding principles for a model to predict new data more reliably), high memory demand, and limited spatial understanding. Attention-based approaches (FocusU-Net, FANet, ResGANet, PraNet, CSAP-UNet), have benefits from dual attention gate (a mechanism for image segmentation) with short-range skip connections, feedback module with hard attention to refine features, lightweight ResNet with modular group attention blocks, parallel partial decoder with reverse attention (a mechanism for image segmentation used in difficult regions), parallel CNN and transformer with self-attention. Challenges include limited fine-grained feature capture, imbalanced local-global context, class imbalance issues and training instability (when the learning process fails or becomes erratic) and overfitting. The advantages of foundation model-based approaches (SAM, SAM2, MedSAM) include: Zero-shot transfer to various tasks, evaluation of SAM for polyp segmentation and capability of universal medical image segmentation. Challenges include poor model interpretability, boundary refinement issues, poor generalization and high cost. Another such program is DeepLabV3+, with an encoder-decoder structure and ResNet architecture. The Kvasir-SEG polyp dataset was used to train the learning-machine, and the model proved high performance metrics such as mean Dice similarity coefficient: 0.9873, mean intersection over union (IoU): 0.9751, making it suitable for polyp segmentation. DeepLabv3+’s performance had higher indicators in polyp segmentation compared to studies conducted with U-Net and similar structures in the assessment of colonoscopic images[ 111 ]. In addition, a distinct model known as DenseNet-169-based online Tigerclaw fuzzy region segmentation, originally developed for dermoscopy imaging, has demonstrated remarkable performance, achieving 98.9% accuracy and high classification robustness for benign lesions, and 97.4% accuracy for malignant lesions, provided that the input data are properly preprocessed and segmented. This model shows strong potential for real-time precision diagnosis of CRC and other malignancies[ 112 ]. Cecal intubation rate (CIR) is also a key quality indicator, as it is associated with a better standardized colonoscopy, better ADR and advanced adenoma detection (ADDR), which is absolutely necessary in screening colonoscopy. Until present, cecal intubation has been a self-acknowledged feature, but AI can help assist colonoscopy and achieve better CIR and ADR by combining the two techniques[ 113 ]. All of the above considered led to lower rates of post-colonoscopy CRC[ 114 , 115 ]. In one study, an AI-based cecum recognition system, known as the AI-common reporting standard, was employed for post-hoc verification during CRC screening colonoscopies. While it did not further improve CIR, it significantly increased both the ADR and ADDR by 5%, with the most notable improvements observed in the proximal colon, the region where post-colonoscopy CRC most commonly arises[ 113 ]. Selvaraj et al [ 6 ] explored a more practical real-time CRC screening AI colonoscopy application with a three-step approach: (1) Automated polyp segmentation using CRPU-Net architecture; (2) Analysis of machine learning for binary classification of segmented regions using scale-invariant feature transform and oriented fast and rotated brief descriptors, followed by feature fusion and dimensionality reduction through principal component analysis to create an optimal feature representation for classification; and (3) Comparison of deep learning exploration for automated classification using state of the art and ViT model. Additionally, they considered the development of a hybrid model colorectal polyp (CRP)-ViT by potentially combining the ViT model with deep learning architectures. As a result, CRPU-Net achieved outstanding segmentation performance with 96.56% accuracy, 95.40% IoU and 97.39% Dice coefficient. Even more, the CRP-ViT model demonstrated promising results in binary classification with 96.59% accuracy, 96.38% sensitivity, 96.80% specificity, 96.82% precision and 96.36% negative predictive value, with the endpoint being the implementation of CRPU-Net and CRP-ViT, AI systems designed for real-time polyp segmentation and potential precancerous polyp classification[ 6 ]. Another way to achieve the quality indicators in CRC screening colonoscopy is using new state-of-the-art 4k image colonoscopes with AI aid computer aided detection endo-aid CADe. Orzeszko et al [ 106 ] examined two groups with 1000 participants each, with one group using CADe. The study found that only polyp detection rate was higher in the CADe group. There was no significant difference in any other segment or parameter[ 116 , 117 ]. Venkatayogi et al [ 118 ] have suggested the use of transfer learning and a vision-based surface tactile sensor (VS-TS) to classify polyps detected during colonoscopy. VS-TS includes a deformable silicone membrane, an optics module, and an array of light-emitting diodes to enhance polyp illumination during imaging. The sensor detects deformation of the silicone when touching polyp phantoms, and each phantom presses against the sensor. In this way, polyps can be identified and characterized according to the detailed textural patterns. Such textural features provided by extraction by the VS-TS can be used and integrated in a SVM algorithm for polyp classification, which requires data from the ImageNet dataset and a preceding training of ResNet-18. Classification occurs on the basis of two versions of ResNet-18: The first uses random weights (ResNet1) and the second, the preceding training on ImageNet (ResNet2). ResNet2 demonstrates an accuracy rate of 91.93%, higher than that of ResNet1, with an accuracy of 54.95%[ 117 , 118 ]. Another network, called area-boundary constraint network, was developed, which focuses on the polyp itself with its characteristics and boundaries, using one encoder for capitalization and two decoders, one for the polyp and the other for the boundary, to raise segmentation effectiveness. It uses selective kernel modules to obtain more accurate and dynamic images of polyps for better segmentation. For training, attentive but diverse network uses 3 databases: EndoScene, Kvasir-SEG and ETIS-Larib[ 119 - 121 ]. A newer way to perform polyp segmentation was proposed by Elkarazle et al [ 122 ], which uses an upgraded version of the multi-scale attention network (MA-NET) aided by a modified Mix-ViT transformer, enhancing its capability to obtain ultrafine-grained visual classification, and therefore more accurate segmentation of difficult polyp types. By adapting the Mix-ViT transformer for this specific application, one can replace the typical convolution-based encoder in MA-NET and leads to better feature extraction at multiple scales. Its advantage is the ability to distinguish between polypoid and non-polypoid regions. The network can enhance features by including a layer that applies contrast-limited adaptive histogram equalization in the CIELAB color for input images and provides a proper solution for segmenting polyps that are small and flat. The necessary training of the model includes Kvasir-SEG and CVC-ClinicDB datasets, with added cross-validation on CVC-ColonDB and ETIS-LaribDB for robustness and applicability[ 122 - 124 ]. One neglected technique of colon visualization is colon capsule endoscopy (CCE). The results were initially disappointing, and it was expensive because of the need to use one capsule for each patient, and each lesion needed a follow-up colonoscopy. During the COVID-19 pandemic, the technique was revisited with better equipment, cheaper and more accurate images, and some promising results were achieved. Researchers are considering coupling it with AI to achieve better polyp matching, which opens up a novel field in endoscopy with potential to address challenges regarding the polyp matching within the same video. Other studies found different figures with more promising uses in CCE. However, all studies found CCE in CRC screening requires some adjustments, including better equipment, lower cost and more input from AI programs[ 125 - 127 ]. Despite the proven potential of AI, refining such techniques is crucial to make CCE more reliable than conventional colonoscopy for lower gastrointestinal diagnostics (Table 3 )[ 128 - 138 ]. The comparison between ileo-colonoscopy and colon capsule endoscopy[ 128 - 138 ] CCE: Colon capsule endoscopy. Considering that screening colonoscopy includes the resection of polyps greater than 1 cm, of which more than 13% are T1 cancers, the question of when to address them via surgery remains. The challenge offers AI an opportunity to create an artificial NN to predict the risk of recurrence using colonoscopy and pathology images. Su et al [ 139 ] found that the AI model prevented more than 34.9% of unnecessary surgical resections compared with United States guidelines in all enrolled patients. Following endoscopic resection of T1 CRC, some procedures require further surgery for lymph node metastasis. An AI model can be employed here as well to reduce unnecessary surgery after endoscopic resection. This possibility has been explored using four classifiers: Regularized LR classifier, RF classifier, CatBoost classifier (CBC), and the voting classifier. Among these, CBC showed the best response based on the Japanese Society for Cancer of the Colon and Rectum guidelines. Furthermore, it demonstrated that by training some AI models, unnecessary surgery can be avoided[ 140 ]. An interesting application explored by Liu et al [ 141 ] is Fovea-UNet, a deep learning model inspired by the human eye fovea. Its aim is to detect and segment lymph node metastases of CRC using CT images. The study used a dataset containing 81 whole slide images (WSI) of metastases from 624 metastatic regions that were extracted and labeled, as well as divided into a training set 57 WSIs (451 metastatic regions) and a test set 24 WSIs (173 metastatic regions). The study directs the model’s focus on the important areas, using a special module adjusted according to feature relevance. In the feature extraction process, they involve a lightweight backbone that is adjusted to a regularization strategy (GhostNet backbone) to cut on demands, while maintaining extraction efficiency. The proposed model has shown its superiority in the segmentation process, promising 79.38% IoU and 88.51% Dice similarity coefficient, better than other models[ 141 ]. When going further and asking AI to look into pathology and predict precancerous and cancerous lesions, unfortunately all of the potential AI engines failed to be more accurate than specialists[ 142 ]. However, AI can perform better in diagnosis using a deep learning model that classifies WSIs into various categories, differentiating the stage of CCR and the normal tissue sample. This model includes a CNN architecture, such as ResNet, which has proven superior in dealing with multiple images, and creates three frameworks: The first uses image level with low resolution to detect potential cancer; the second uses cellular level images that employ the CNN to differentiate cancer from noncancer; and the third combines the first two for better diagnosis, reaching 94.6% accuracy on the Cancer Genome Atlas (images from Cancer Genome Atlas) dataset and 92.0% accuracy on an external dataset with real life images from hospitals[ 143 ]. To better train AI models, some researchers have provided histological images of CRC samples from a dataset available by Kather with 5000 tissue tiles, to help distinguish between tumor and stroma and integrate the data through a transfer-learning-based binary classifier. Such framework employs four different CNN architectures: Visual geometry group (VGG) 19, EfficientNetB1, InceptionResNetV2 and DenseNet121, with background information from ImageNet dataset. Tuning the classifier on specific features of the CRC dataset, one can also retain the learned features. The results of these architectures are of higher accuracy: 96.4%, 96.87%, and 97.65%, for VGG19, EfficientNetB1 and InceptionResNetV2 respectively, which go beyond the presented reference values of the study. In conclusion, applying transfer-learning using pre-trained CNNs works as a booster for classifying tumor and stroma regions in terms of histological CRC images[ 144 , 145 ]. One can go further towards feature extraction and use optimal deep feature fusion approach on biomedical images method, median filtering to eliminate noise, and a combination of three deep learning models: MobileNet, SqueezeNet and SE-ResNet. A selection of hyperparameters can be performed using the osprey optimization algorithm; finally, a deep belief network model is engaged to better classify CRC. The technique was tested on the Warwick-QU dataset and achieved 99.39%accuracy[ 146 ]. In the process of CRC screening, imaging of the colon (computer-tomography and magnetic resonance imaging) is also useful, especially for patients who are ineligible for colonoscopy (if there are fixations that cannot be passed or if anesthesia is not recommended), or in the future, when more noninvasive techniques for screening are expected to be developed. Even in this field, the results are relatively subjective and tend to be operator-dependent. Under such circumstances, the AI intervention may optimize diagnosis, proven by some studies that use RK-Net with unsupervised learning (UL) techniques and deep learning. The network takes cluster images, performs extraction using Mo-bileNetV2 and learns to identify cancerous tissue. The benefits of using UL are reduced training time by selecting only the most relevant clusters, reducing costs and time for training, while achieving 95% accuracy[ 147 ].

Since their invention, computers have evolved into diverse branches and become integral to nearly all scientific, engineering and healthcare fields. They enable the development of programs designed to replicate aspects of human intelligence and behavior. In healthcare, AI has emerged as a powerful tool, particularly in the prevention, diagnosis and management of diseases, including CRC. AI operates through two pathways: Virtual and physical. The virtual pathway encompasses machine learning and deep learning, with the latter representing a specialized subset of the former. Machine learning involves processing input data to identify patterns using statistical and analytical techniques, without the need for explicit programming[ 37 ]. In unsupervised machine learning, the program autonomously identifies patterns within data using statistical methods, without predefined labels or outcomes. In contrast, supervised learning relies on known outputs to guide the training process toward specific goals. Deep learning employs artificial neural networks to model complex relationships. It includes DNNs for feed-forward processing, recurrent neural networks for sequential data analysis, and convolutional neural networks (CNNs) for interpreting grid-like structures. These architectures are particularly well-suited for tasks such as biomedical signal processing, feature learning, and medical image recognition[ 38 - 44 ]. Recent advancements in AI have led to the development of soft AI colonoscopy, designed to enhance both lesion recognition through computer-aided detection (CADe), and optical diagnosis through computer-aided diagnosis (CADx). These AI-powered colonoscopy systems employ deep learning algorithms to analyze real-time endoscopic images, thereby improving detection rates for adenomas, serrated lesions and CRC. Studies have shown that AI-assisted colonoscopy improves adenoma detection enabling earlier intervention and leading to better patient outcomes. The benefits are particularly notable for less-experienced clinicians, as the detection rates approach those of experts. Moreover, AI integration can support training and education, promote standardized procedures, and enhance the consistency and reliability of screenings across healthcare settings, thereby improving accuracy and efficiency. As AI technology continues to advance, its application in CRC screening has the potential to revolutionize early detection, reduce CRC-related mortality, and reshape the future of cancer prevention and care[ 41 ]. Despite these benefits, challenges such as the cost of AI implementation, data privacy concerns and the need for extensive clinical validation must be addressed before widespread use. Studies indicate that the application of AI in healthcare attracted approximately 6.6 billion USD in investments from public and commercial sectors in 2021. In the United States alone, the use of AI is projected to generate annual savings of up to 150 billion USD by 2026, reflecting its growing impact on healthcare efficiency and cost reduction[ 11 ]. In this review, we present the latest advances in the field, along with with future applications and expected limitations in CRC screening.

In 2025, ChatGPT (version 4.5) was evaluated for geographic and linguistic accuracy to provide recommendations for CRC screening and monitoring. Due to detected variations across countries and languages, the AI model requires further validation according to each context. Large language models need to be equipped with high-quality clinical data, sensitive to the region they originate in, so that they can be used in real-world settings in a reliable manner. The characteristics they still lack are standardization and transparency. All these could be enhanced by reducing present limitations in data sourcing, time and version differences and title-recommendation discrepancies. Healthcare professionals should ponder all input when using AI-generated advice in clinical practice[ 154 ]. It has been demonstrated that deep learning, radiomics and multimodal data integration may achieve similar results as expert endoscopists in matters of endoscopic image examinations. However, the challenges that arise are due to limited model generalization (fragmented datasets), algorithmic limitations in rare conditions, insufficient training data and ethical issues. To overcome these, multicenter databases should be acquired, AI tools in prospective trials should be validated, and programs to raise clinical trust, unified ethical/regulatory frameworks should be developed[ 155 ]. AI models can also assist polyp classification in computed tomography colonography, aid in and select the referral of patients for endoscopy (as a second reader). In terms of economic impact, such aid would result in cost-effectiveness. According to guidelines, the recommendations are the following: The resection of CRPs ≥ 10 mm and CRPs of 6-9 mm depending on age or other conditions should be followed by endoscopy, while polyps less than ≤ 5 mm should undergo screening computed tomography colonography. However, these recommendations have proven to not be cost-effective, only gaining 464407 United States dollar per life-year, compared to 59015 United States dollar for polyps with a size of 6-9 mm and 151 United States dollar cost savings for each patient, for polyps larger than 10 mm[ 156 - 160 ]. AI provides a benefit as a second reader for polyp classification, turning endoscopic referral after computed tomography colonography into a more effective tool when analyzing and diagnosing adenomatous and non-adenomatous CRPs[ 161 ]. Furthermore, CADe using computed tomography colonography data could improve sensitivity by analyzing images obtained in two positions. At the same time, both sensibility and sensitivity can be increased by a deep learning algorithm that integrates training to detect small and large polyps, low-dose pretreatment, non-uniform intestinal dilatation, or low-dose imaging[ 162 , 163 ]. An intriguing recent study describes how four AI models integrated with ColonView test results and clinical features of the patients have outperformed the diagnostic accuracy of the conventional LR method in FIT-based screening. These models are: SVM, NN, RF and GBM and confirm the augmentation of the diagnostic accuracy rate when combined with non-AI/clinically obtained data[ 164 ]. Increasing ADR and polypectomy rate in AI-assisted colonoscopy is a goal for all gastroenterologists worldwide. In Asia, the use of AI-aided colonoscopy (real life) has brought about a significant difference for endoscopists, raising ADR from 24.3% to 30.4% and the polypectomy rate from 28.4% to 33.6%[ 165 ]. Finally, the tendency for such studies is also remarked regionally in small countries in certain regions; an example is Romania: South-West Oltenia studied and suggested a refined framework for CRC risk stratification, using NN-based composite scoring for stratification of FIT[ 166 ].

Beyond all scientific fields of CRC screening in which AI has been involved, rather complex and sometimes hard to implement, the real-life clinical impact should encourage further research and innovative ideas. In matters of clinical aspect, AI has increased the ADR, to reduce missed lesion rates, by extending the spectrum of colors, areas that are difficult to visualize, flat polyps that are hard to identify. Moreover, it improves in vivo polyp characterization by increasing the prediction rate for the stage of polyp/cancer, solidifying the link to histopathology, better identifying hyperplastic polyps to resect and not collect (to reduce the number of unnecessary polypectomies), predicting bleeding in performing a polypectomy, and in the end stratifying the risk for every individual patient. Another benefit is that it standardizes colonoscopy quality by reducing inter-individual variability, raising examination quality, aligning the description language (for both colonoscopy and histopathology), and raising colonoscopy quality for beginners. Finally, it integrates and promotes CRC screening by identifying the population who needs screening, and highlighting the risk population and raising awareness. It may also reduce the rate of duplication or omission and correct the weak points of screening. It also has a significant economic impact by reducing the cost of double colonoscopy when bowel preparation is weak, reducing unnecessary histopathological assessment of polypectomies in patients who do not require it, and raising the effectiveness of the procedures, shortening their duration, and possibly reducing their number or the interval between re-screening. The list continues, with a reduction of the number of unnecessary surgeries, prioritization of cases that need emergency colonoscopy or additional attention from an expert, rather than a beginner. In the end, reducing the costs of advanced CRC by detecting early lesions, and either standardizing the treatment in each patient case or adjusting to its stage should not be neglected.

AI inevitably remains an intriguing field due to its variability, training capacity and impact on various medical societies across the world. Table 4 summarizes the guidelines provided by several important associations in terms of CRC screening. Real-life recommendations of world-wide major associations[ 148 - 153 ] AI: Artificial intelligence; FIT: Fecal immunochemical test; CT: Computed tomography; CADe: Computer-aided detection; FDA: Food and Drug Administration. Recommendation (with grade: No recommendation-very low certainty of evidence): In adults undergoing colonoscopy, the American Gastroenterological Association (AGA) makes no recommendation on the use of CADe-assisted colonoscopy[ 148 ]. Currently, there are draft guidelines that suggest CADe as a quality-enhancing tool, especially in polyp finding in adults[ 148 ]. Remarks[ 148 ]: CADe is already acknowledged by the AGA to improve ADR, but overall, there is a very low certainty of evidence on long-term outcomes for decision making in CRC incidence and mortality[ 148 ]. An update is planned for this recommendation because CADe continues to improve as an AI application[ 148 ]. There are evidence gaps on data from studies referring to the impact on advanced adenomas, difficult-to-detect adenomas, access to colonoscopy, patient values and preferences, cost-effectiveness or impact on long term outcomes[ 148 ]. The combination between CADe and computer-aided diagnostics may add to post-polypectomy surveillance in the context of this technology, and further studies may be applied on diagnosing and resecting polyps with CADe alone[ 148 ]. One of the most important issues is the concern for security in patient privacy, data security and ownership of AI platforms used in healthcare. According to some authors, arising healthcare errors should be a responsibility shared between the doctors who used the device and the designers[ 148 ]. There are several CADe systems on the healthcare market with periodic updates, but these should be more transparent and accompanied by training tools and facilities. Even for educational purposes, there are authors who believe it would negatively impact skill development[ 148 ]. Current advances in AI allow for the development of AI-based algorithms that can be applied to endoscopy to augment endoscopist performance in lesion detection and characterization. This is equivalent to subjectivity reduction in reporting quality and helping to build smart endoscopy algorithms, which can subsequently optimize workflow in endoscopy, including improved documentation[ 149 ]. Using AI and machine learning aids in predictive models, diagnosis, and prognosis, but high-quality data with multidimensionality are needed for risk prediction and forecasting of specific clinical conditions and their outcomes, whenever using machine learning methods[ 149 ]. Large data pool and cloud-based tools are essential for the clinical research in gastroenterology to go a step further. To understand maximal extent of the diseases and make treatment decisions, multimodal data are required for training AI[ 149 ]. Understanding how to evaluate AI algorithms in the gastroenterology literature and clinical trials is important for gastroenterologists, trainees, and researchers, therefore implying educational efforts to be made by gastroenterology societies[ 149 ]. AI algorithms have to be transparent, easy to interpret and explain, so they can be adopted in the clinical practice of gastrointestinal endoscopy. Regarding the payment of AI in endoscopy, a thorough evaluation of AI systems needs to be performed to purchase the most reliable and cost-effective endoscopy guide. Studies to guide reimbursement may also be needed[ 149 ]. Evidence standards for AI in gastroenterology are not yet well-defined and need further development. Several aspects still remain to be dealt with and require a balanced view of AI technologies and an active collaboration between the medical team and staff from technology, industry, computer science fields, last but not least researchers. This is critical for a suitable advancement of AI in gastroenterology[ 149 ]. The general belief of this panel is that patients would opt-in for CADe assistance for screening or surveillance colonoscopy, during their procedure, on condition that they are adequately informed. They should be made aware of both potential benefits and limitations, in the case of such measure for reducing CRC incidence and mortality[ 150 ]. However, this recommendation is weak because it bears a significant uncertainty of evidence, low absolute benefits for CRC incidence and mortality, and it increases the patient burden associated with CADe use, especially by over-diagnosing polyps and the need for more surveillance by colonoscopy[ 150 ]. Medical innovation is often most exciting in its early stages, particularly during the design and promotion of preliminary products. However, before new medical devices can be implemented in clinical practice, they must undergo extensive bureaucratic processes that are both time-consuming and expensive. However, these challenges should not discourage the development of brilliant ideas in medicine. To maintain high standards and strict regulations, the regulatory authorities should be involved from the start to find solutions. For example, in Japan, the Pharmaceuticals and Medical Devices Agency, Medical Device Agency and the Ministry of Health, Labor and Welfare provide opportunities for consulting in matters of regulatory approval and reimbursement applications. Several public bodies, such as Japan Agency for Medical Research and Development, offer support, including financial support, in the whole process of implementing new medical devices and applications, with an even greater opening towards challenging projects[ 151 ]. Finally, even after achieving regulatory approval and reimbursement, the journey is not over: Any new medical or AI device should be continuously monitored for benefits and harms that potentially arise during use, and this also requires a tremendous effort, all for the optimization of healthcare[ 151 ]. In 2025, the data are promising, but in a real medical setting, there is one question to be answered regarding the AI assistance, before implementation: “Who is ultimately responsible for a clinical result?” Even with its continuous improvement and refinement, AI cannot avoid false-positive or false-negative diagnosis, thus human doctors must take responsibility for the final clinical decision. The Japan Gastroenterological Endoscopy Society provides guidelines for the clinical use of AI software ( https://www.jges.net/wp-content/uploads/2023/05/AIsoft.pdf ). The guidelines clearly state that physicians are responsible for diagnosis, and that AI should only be used adjunctively to enhance performance in various clinical settings and contribute to the better health of people around the world. Japan’s first major AI-colonoscopy system, EndoBRAIN, received regulatory approval and reimbursement starting in 2024[ 152 ]. In conclusion, Japan, as a leader in the field of gastrointestinal endoscopy, takes on the responsibility to demonstrate the clinical and economic benefits of AI in gastro-intestinal endoscopic diagnosis, to initiate social implementation and to build further evidence[ 152 ]. In January 2021, the Food and Drug Administration (FDA) launched its AI/machine learning based software as a medical device action plan, which sets the foundation for regulating AI in healthcare. Its main points are: A tailored regulatory framework for such software, adherence to good machine learning practices, patient-centered transparency, algorithmic bias control, robust performance and real-world performance monitoring. The AI tools for CRC detection cleared by the FDA are: Gastrointestinal genius (which assists real time colonoscopy interpretation with a raise in ADR from 42% to 55.1%); SKOUT system (a CADe device for trained gastroenterologists to flag potential polyps in real time colonoscopy); Olympus CADDIE (the first cloud-based AI platform for CRC colonoscopy assistance allowing remote updating)[ 153 ].

AI has the potential to significantly transform CRC screening at multiple stages of the patient diagnostic process. Its integration begins with improving diagnostic accuracy of noninvasive tests, such as FIT, by applying algorithms that optimize threshold settings and patient stratification. In addition, AI-assisted tools can support bowel preparation, guiding patients through personalized instructions to ensure high-quality colonoscopy outcomes. During endoscopic evaluation, AI has the greatest impact by increasing ADR through real-time image analysis, better characterizing detected polyps and reducing the rate of missing precancerous lesions. It can also reduce the number of unnecessary interventions endoscopic and surgical procedures. Beyond detection, AI contributes to therapeutic decision-making in the post-screening stages, helping in surgical planning and lymph node removal, enhancing both safety and outcomes by reducing unnecessary interventions. Other technologies that are less user-friendly, such as colonic capsule endoscopy, offer a promising alternative screening approach when combined with AI, especially in populations with limited access to traditional colonoscopy. In imaging, AI improves the accuracy of interpretation, particularly in computed tomography colonography and capsule endoscopy, thereby expanding screening capabilities beyond conventional methods. In addition, AI facilitates standardization of histopathological analysis, reducing interobserver variability and improving the consistency of CRC diagnostics. The development of AI-based applications and networks can streamline the entire CRC screening process, optimizing workflows, ensuring compliance with follow-up and facilitating data sharing for population-wide screening programs. From a broader perspective, the implementation of AI in CRC screening promises positive clinical and economic impact, with the potential for earlier detection, reduced interval cancers, and cost savings due to more efficient use of resources. There are, however, limitations that remain, including the generalization of algorithms, data privacy concerns, regulatory hurdles and the need for extensive clinical validation. In conclusion, AI holds the promise of substantially enhancing the effectiveness, efficiency and equity of CRC screening, but its full potential will only be achieved through careful integration, rigorous evaluations and ethical consideration.

This review has primarily focused on the benefits of AI assistance during various steps of CRC screening, diagnosis and medical conduct. However, there are also limitations that should be mentioned. First, there is no perfect technology, especially in the field of medicine. Deep learning mechanisms perform well given that they are provided sufficient data, such as thousands of images with polyps, histopathology, computed tomography and magnetic resonance imaging image slides. Considering that even in real life cases, polyp classification and recurrence pattern prediction are still challenges, even for experts, slight differences are likely to arise in deep learning. While expert opinions can be accompanied by arguments, AI may fail to provide reasonable explanation. Another aspect to be considered is population variability due to age, gender, and race. In this case, AI algorithms can have biased results; for example, NNs can perform well for a certain population (for which the program was trained) but make wrong decisions for another (for which the data input is low). Besides population, there is significant variability between endoscopes that can alter and bias the algorithms. Also, results can be influenced by situations when the human eye proves its superiority to the AI due to its fast adaptability. This is the case when AI cannot discriminate and analyze correctly due to endoscopic factors such as peristaltic movements, gas bubbles, light variations, contrast, leading to lower performance. Last but not least, there are still economic and ethical perspective problems to be addressed: The responsibility of AI errors (for example, missed cancerous lesions), patient data confidentiality, workforce reduction in the case of AI overtaking, and the financial burden. Patients have become increasingly preoccupied by the confidentiality of their personal information, every medical procedure requires a signed consent; involving computer programs raises new worries among patients the fear that their personal data may virtually leak worldwide. If AI is to be officially involved in diagnostic procedures, the financial expenses may initially increase, as most programs to be implemented come with high costs. Therefore, the questions are: What healthcare department decides which programs and devices should be implemented in the patient workflow; how they can be made available worldwide, who should pay for them, and what happens in the scenario in which human task force is replaced by AI?

Cancer represents a major global health burden, affecting social, clinical, psychological and economic domains, and demands the mobilization of every available resource to address it effectively. According to the Agency for Research on Cancer, there were 20 million new cases of cancer per year in 2022, with over 9.7 million cancer-related deaths per year. Colorectal cancer (CRC) is a major cause of cancer-related death worldwide, requiring early and effective CRC screening to improve survival rates. Worldwide, CRC ranks third in incidence, accounting for 1926425 new cases (9.6%) across both sexes, 1069446 in males and 856979 in females, according to the most recent Global Cancer Statistics report (2022). In terms of mortality, CRC ranks as the second leading cause of cancer-related death worldwide, accounting for approximately 904019 deaths annually (9.3%)[ 1 ]. In the United States, the burden is similarly high, with CRC representing nearly 10% of all cancer cases and responsible for 8%-9% of cancer- related deaths[ 1 - 3 ]. According to the 2022 Global Cancer Observatory statistics, in the United States, breast cancer ranks first in incidence (12%), followed by lung cancer (11%) and CRC (10%). However, in terms of mortality, CRC ranks second (9%), after lung cancer (18%)[ 1 , 4 - 6 ]. The Indian Council of Medical Research reports annual incidence rates (AIRs) for colon and rectal cancer in male patients of 4.4 and 4.1 per 100000, respectively. In comparison, the AIR for colon cancer in female patients is approximately 3.9 per 100000[ 7 ]. These data have prompted countries worldwide to implement improved screening strategies. In the European Union, the target is to achieve screening coverage of 90% among the eligible population (ages 50-74), using the fecal immunochemical test (FIT) test as the initial method, followed by colonoscopy for positive cases. In the United States, recommendations differ slightly: Individuals at average risk (those aged 45 and older) are advised to undergo either a stool-based test [FIT, fecal occult blood test (FOBT) or DNA-based test] or a visual examination (colonoscopy, rectosigmoidoscopy or computed tomography colonography). Those at higher risk should begin screening earlier and use more specific diagnostic methods[ 1 ]. Traditional CRC screening methods include non-invasive stool-based tests, such as the guaiac FOBT, FIT, and multi-targeted stool DNA test, as well as more accurate diagnostic procedures like colonoscopy. Whether used as a primary screening tool or as a follow-up after a positive FIT, colonoscopy remains the gold standard for detecting and removing precancerous polyps and identifying early-stage cancers. However, its effectiveness can be affected by factors such as operator skill, patient variability and lesion visibility. The persistence of missed lesions underscores the need for more efficient and accurate screening techniques[ 8 ]. Studies have shown that up to 22% of polyps may be missed during CRC screening colonoscopies, and approximately 8% of cancers develop within 3 years following a screening colonoscopy[ 9 , 10 ]. Studies indicate that several biomarkers are already in use for diagnosing CRC; however, their specificity for CRC cells remains limited. These include carbohydrate antigen (CA) 19-9, a tetrasaccharide expressed on the surface of pancreatic and colonic adenocarcinoma cells, and CA125, which is detected in both colonic and ovarian cancers and has also been associated with endometriosis. DNA-based molecular biomarkers function by detecting methylated DNA in stool samples. Improved diagnostic accuracy can be achieved by combining the FIT with a multi-target stool DNA test to detect and quantify hypermethylated DNA. Stool DNA testing demonstrates higher sensitivity than FIT alone (92.3% vs 73.8%); however, it is also associated with a higher rate of false positives. Another serum-based blood test developed and studied in the United States is Epi proColon, which uses a real-time polymerase chain reaction (PCR) technique to detect methylated SEPT9 , a biomarker associated with CRC. However, its sensitivity and specificity are relatively low: 68.2% and 79.1% respectively, regardless of cancer stage. Due to its higher rate of false positives, Epi proColon performs less effectively than FIT and colonoscopy. Other methylation markers that have been investigated include SFRO2 , VIM , FBN2 , TCERG1 and serum vimentin methylation. Their diagnostic performance for early-stage CRC detection generally falls within a moderate range of specificity and sensitivity. In addition to biomarkers, circulating tumor cells (CTC) have been studied, analyzed and demonstrated promise in cancer detection. However, this technique is limited by the low concentration of CTCs in the blood (1 CTC per 10 9 red blood cells) and by variability in sampling methods. The efficiency of CTC capture and enrichment largely depends on their physical and biological properties, including density, size, deformability, electrophoresis and affinity-based capture. Common selection approaches involve identifying cells that are CD45-negative and EPCAM-positive[ 11 - 23 ]. Another promising biomarker under investigation is tumor-associated circulating transcripts. These RNA-based liquid biopsies leverage artificial intelligence (AI) models as generalized linear models, random forest (RF), gradient boosting machine (GBM), deep neural networks (DNN), and automated machine learning. Among these, the DNN model has demonstrated superior performance, achieving optimal sensitivity (85.7%) and specificity (90.9%) for the detection of early-stage CRC. Of the ten markers studied, EPCAM , NPTN , KRT19 , MKI67 and VIM were found to be highly expressed in CRC cell lines, whereas FOXA2 and MCAM showed lower expression levels. In contrast, ERBB2 , TERT and SNAI2 were undetectable[ 24 - 26 ]. SEPTIN9 is another biomarker found to be useful in more advanced pathological stages of CRC, as levels of methylated SEPTIN9 in peripheral blood increase with disease progression. Its reported sensitivity for detecting CRC ranges from 75% to 79.3%, whereas its sensitivity for identifying adenomas remains suboptimal at approximately 27%[ 27 , 28 ]. In recent years, the gut microbiota has been intensely studied as a key to early CRC and adenoma detection. The gut microbiota can influence several processes, including metabolic regulation, inflammation control, and epigenetic reprogramming. There is increasing evidence that the gut microbiota can be leveraged to identify individuals at high risk for CRC and early detection. After performing 16S rRNA sequencing on fecal samples from individuals with positive FIT, researchers found a significant increase in the abundance of Proteobacteria in those with colonic lesions compared with healthy controls. Another study identified elevated levels of Fusobacterium nucleatum in the oral cavity, tumor tissue, and feces of individuals with colonic lesions. The presence of this bacterium demonstrated a sensitivity of 82% and specificity of 62% for distinguishing colorectal adenomatous polyposis from healthy individuals, and a sensitivity of 66% with specificity of 90% for differentiating adenomas from CRC[ 29 - 35 ]. The detection of free fatty acids (FFAs) in serum represents another promising approach for early CRC diagnosis, as CRC patients exhibit significantly lower FFA levels compared to healthy individuals. Combined analysis of FFAs such as C16:1, C18:3, C18:2, C18:1, C20:4, and C22:6 has achieved a sensitivity of up to 84.6% and specificity of 89.8% for early CRC detection. Furthermore, the combination of C16:1, C18:3 and C18:2 demonstrated a sensitivity of 70% and a specificity of 81% in distinguishing benign intestinal diseases from CRC[ 36 ].