Medical needs related to the endoscopic technology and colonoscopy for colorectal cancer diagnosis

To identify medical needs, a 2-step process has been followed. Initially, a reduced number of endoscopists were interviewed to extract a draft list of needs, which was later confirmed and improved by surveying a broader number of clinicians.

This study exempted from review by the Institutional Review board because it was not within the scope of Law 14/2007 of 3rd July on Biomedical Research. This national regulation indicates that human health-related research involving invasive procedures needs approval by the ethical committee, but in this study the participants were not involved in invasive procedures. For the same reason, it was not considered necessary for participants to sign a written informed consent. Only verbal informed consent was considered sufficient. All participants were informed about the objectives, length and procedure of the study, the institutions responsible for the research, the contact people, the confidentiality and anonymity of the research data, and that their participation was voluntary, and they could withdraw from the study at any time if they wished to.

After analysing the survey results, no significant differences have been found in either of the three considered study groups. This analysis suggests that existing medical needs would be equally perceived by all endoscopists, and would be not biased by belonging to one group. However, this should be considered with caution as qualified endoscopists with special interest in cancer screening and/or therapeutic endoscopy and expert endoscopists are overrepresented. Therefore, only global results from interviews and surveys of all participants are discussed, from which the medical needs regarding the colonoscopy procedure and the used technology can be extracted and divided in three different fields: clinical needs, Computer Aided Diagnosis (CAD) system needs and operational/physical needs. Verbatim quotes from participants are provided in Additional file 3.

From a clinical point of view, the main concern for endoscopists is to improve polyp detection, especially for flat polyps, to reduce the polyp missing rate. Once polyps have been detected, the next identified need is to minimize the resection of hyperplastic polyps following the “diagnose and leave” strategy. Furthermore, this concern leads to the clinical need related to in-situ classification of detected polyps, so that either the “resect and discard” or “diagnose and leave” strategies can be followed if deemed necessary. This way, costs associated to extraction, processing and analysis of hyperplastic polyps could be reduced, as well as the time for diagnosis, minimizing negative psychological impact in patients, as stated by an endoscopist. Clinicians also need to reduce recurrence rate by assessing lesion margins and inspecting the remaining tissue, avoiding re-interventions which negatively affect the patient prognosis and increase the risk of complications (such as bleeding, perforation, etc.). Equally important is to assess the depth of submucosal invasion for detected lesions, as it is an essential parameter to decide whether endoscopically resect early colorectal cancerous lesions.

Survey results (Table 1) corroborate these above-mentioned needs from the interviews. Endoscopists demanded tools to improve the polyp detection. They felt that current methods for detection are insufficient and therefore that current approaches could be improved by using advanced technologies for polyp detection and characterisation. The rapid identification and classification of polyps was considered as clinically useful. In fact, 88.0% of respondents indicated that they would like to have an automated system that would suggest the diagnosis without having to make an interpretation of the lesion.

The overall preference was a tool to detect flat or depressed polyps, as 89.7% of endoscopists indicated that this assistance would be more beneficial for this type of lesions, reflecting the consensus that they are more difficult to detect than elevated lesions. The most desired characteristic to be provided by such system is the infiltration probability (83.3%), followed by the estimation of the polyp depth (71.4%), a diagnostic suggestion (65.5%) and polyp margins (63.3%). All these features would assist on deciding how to proceed and potentially avoid unnecessary biopsy or polypectomy if demonstrated to be of adequate accuracy.

According to the results from the interviews, endoscopists expect the software, or computer aided diagnosis (CAD) systems, to support the polyp detection and diagnosis with additional information for a better diagnosis. They also demand that the CAD system provide them with information about physical characteristics of the polyps, such as the size, and its location. Besides including polyp detection and characterization in the CAD system, it is also important to accurately visualize the polyp contour for safe margin resection and avoiding thus recurrence rates. Clinicians stated that a CAD system that indicates polyp invasiveness or remaining adenomatous tissue left after resection is needed, as it is currently not available. Finally, for a better understanding and a user-friendly environment, all this information provided by the CAD system is requested to be mainly visual.

Survey also shows these trends (Table 2), as most endoscopists prefer that the CAD system provides feedback as visual cues to show polyp detection directly highlighting the polyp on the endoscopic image (also called augmented display) (72.7%) instead of being shown on the screen away from it (43.4%). Regarding polyp diagnosis suggestion (classification), a high proportion of endoscopists (83.3%) preferred a ‘traffic light’ display, with different colours representing different histological predictions, rather than text (31.0%) or an audible alarm (3.6%).

On the other hand, the most helpful method to establish the malignancy of polyps is the Kudo’s pit pattern (47.8%), over the Paris classification (23.3%) or the vascularity (12.9%). Therefore, future developments should consider presenting information in this manner, although more than half of the participants (65.9%) think that the existing methods to establish the malignancy of a lesion are not simple or reproducible enough.

Interviews with panels of expert endoscopists gave also insights into the operational needs. Firstly, the morphology of the colon itself makes it difficult to visualize and detect certain polyps. For that, they demanded better lighting inside the colon to perform the colonoscopy, shortening thus the time spent in reviewing or cleaning dark areas. To meet the need for a better visualization and exploration of the mucosa, the use of flexible/rotatory tips in the endoscope should be further explored. This way, polyps located in hidden areas of the colon could be detected by just rotating the tip. However, the flexibility of the instruments to be inserted is also a concern for endoscopists, as they should be flexible enough to pass through the working channel, but rigid enough to transmit the movements and force from outside. Finally, endoscopists would like to use equipment with better image quality.

The interviews to analyse consist of a collection of six documents, one for each interviewed endoscopist. After applying the LDA method, two topics that best describe this collection were found. For each topic, the words distributions were represented as word cloud charts, so the larger the occurrence of a word is, the bigger and bolder it appears in the chart. Details of the LDA results are provided in Additional file 4.

The first topic revolves around the patient (counted 22 times). This might be because some technical problems during colonoscopy are mainly related to the patient. While some can be easily overcome, such as the poor preparation of the colon before the colonoscopy, others are intrinsic to the patient, such as the anatomy of the colon in terms of folds or haustrum. In any case, they complicate to see all polyps and dark zones, as well as the progress of the endoscope. Moreover, since colonoscopy has to be performed in a limited time, these patient-related problems also reduce the exploration time. Lastly, reducing the time until the diagnosis has not much impact on the patient’s prognosis but rather an executive/management impact.

In the second topic, polyp stands out over the rest of words (counted 36 times), suggesting that the focus is placed on correctly knowing how to detect and identify the polyp based on endoscopy imaging, as supported by other words in the topic, such as the ones equivalent to polyp, like lesion, injury, or adenomatous, other related to placements of polyp, like position, area, or colon, or actions about the polyp, like measure, detect, locate, or identify.

For better in-situ polyp detection and classification, as demanded by clinicians, various technologies are available to enhance the endoscopic image. Firstly, the NBI highlights the contrast between vascular structures and the surrounding mucosa [30], while the Flexible spectral Imaging Color Enhancement (FICE) technique, and its extension Blue Laser Imaging, consider morphological features, pit pattern and vessel characteristics of polyps [31, 32]. On a different approach, the linked color imaging finds pixel color differences in neoplastic and healthy tissue [33]. Lastly, the iScan provides real time virtual chromoendoscopy for a detailed view of the mucosal and vascular patterns [34].

Besides these techniques, the standard polyp detection with white light colonoscopy can be improved further by means of fluorescence [35, 36] or autofluorescence [31] imaging. This last approach does not use exogenous agents employing both visible [37] and near-infrared [38] light. In contrast to this, endocytoscopy is a high-resolution microendoscopy technology that detects fluorescence after the systemic or topical administration of an exogenous fluorophore [39].

The emerging advanced optical imaging technologies, such as confocal fluorescence microscopy (CFM), optical coherence tomography (OCT), Raman spectroscopy (RS), hyperspectral spectroscopy or multi-photon tomography (MPT), show great potential for assisting clinicians in the early detection of cancerous diseases [40]. CFM provides real-time microscopic images underneath the surface mucosa of 40–70 μm depth, but requires expensive equipment, significant technical expertise and exogenous fluorophore. In case of OCT, this technology provides high-resolution assessment of subsurface structures without contrast agent. However, there might be a delay in image acquisition and in vivo studies in humans are still lacking. RS details the tissue composition at a microscopic level and spectral information is suitable for automated analysis, but further characterization in humans and the development of endoscopic platforms is required. Finally, MPT high-resolution images show histologic features and changes in tissue without contrast agent, but further characterization in humans and development of endoscopic platforms are missing. Thus, combining imaging technologies that offer complementary information so flaws can be compensated is at stage on the current development of new diagnostic solutions [31].

The above-mentioned enhanced imaging techniques have been commonly applied to identify and characterize colonic lesions but they are less focused on identifying incomplete excision and residual tissue during polypectomy or recurrent disease on follow up, even though there is some evidence of its potential [41]. As for NBI, its effectiveness has not been shown to determine complete excision of diminutive polyps (< 5 mm) [42], but the detection of recurrence when examining previous scars was enhanced when combined with high definition (HD) white light [43]. On the other hand, chromo-endoscopy with 0.13% indigo carmine has showed a significant increase in residual polyp detection compared to alone while light endoscopy in predominantly larger polyps (> 5 mm) [44]. Alternatively, the Fourier transform infrared spectroscopy, that provides chemical analysis at the molecular level, helps determining whether the resection margin contains foci with recurrence potential [45].

Finally, regarding the clinical need to estimate the depth of submucosal infiltration of polyps, the endoscopic ultrasonography (EUS) and narrow band imaging-magnifying endoscopy (NBI-ME) are often used. The axial images of the tumors produced by EUS could detect the invasion of the normal layered structure of the colorectal wall [46], while the NBI-ME combination could provide visualization of capillary vessels and their fine structure in the surface layer [47, 48]. A comparative study has shown that NBI-ME is more appropriate for estimating the early CRC invasion depth before treatment to avoid unnecessary surgery, whereas EUS should be applied to early CRC in which the decision to conduct endoscopic resection is difficult [49]. Furthermore, the probe-based confocal laser endo-microscopy is also useful for the differentiation of normal submucosa from carcinoma infiltration, especially when is accompanied by severe fibrosis. This technique provides more sensitivity and accuracy than magnifying chromo-endoscopy, but large-scale prospective studies are needed to further evaluate the clinical impact of its use during endoscopy [50].

The imaging technologies should be complemented with advanced image processing methods that facilitate the detection, analysis and diagnosis of polyps on real time through a CAD system.

On the one hand, many detection methods based on artificial intelligence are already available, although clinical evaluation is still missing [51]. ENDOANGEL helps improving the ADR by identifying endoscope slipping and previously seen frames, thus avoiding blind spots [52]. Similarly, the Automated Polyp Detection Software (APDS) [53] analyses colours, structure, texture and motion information of polyps, and marks a region of interest with small green rings on the endoscopic image. This visualization method matches the preferred option in this study. However, APDS should be further improved before its actual clinical application as it did not detect more polyps than those already detected by clinicians [54]. Wang et al. [55] present an automated polyp detection system which was used during colonoscopy providing as output simultaneous visual and sound alarms. This system increased the ADR, predominantly thanks to the detection of hyperplastic polyps and diminutive adenomas, but there was no difference in the detection of advanced adenoma or sessile serrated lesions [53]. Finally, the GI-Genius system shows a faster reaction time for polyp detection than endoscopists while increasing also the ADR [56]. Nevertheless, there is still room for improvement and methods based on deep learning and/or machine learning are still being developed to explore different approaches, such as the extraction of fractal dimension of wireless capsule endoscopy images used for the identification of abnormal frames [57], or the use of a weakly supervised convolutional neural network that can be trained solely with semantically annotated images, indicating whether they contain anomalies or not [58].

Efforts have also been focused on CAD systems for the classification and diagnosis of colonic polyps. EC-CAD performs an analysis of nuclei and texture from endocytoscopic images of colonic tissue to estimate the probability of non-neoplasm, adenoma and invasive cancer [59] and on the same line, but using endocytoscopy with NBI, a CAD system evaluates cellular, glandular, and vessel structures of colonic polyps and provide the probability of neoplastic or non-neoplastic diagnosis [60]. Similarly, an artificial intelligence model for real-time characterization of colorectal polyps has been developed based on NBI video frames [61]. This system grades the polyp according to the NICE classification into type 1 – hyperplastic or type 2 – adenomatous and provides the associated probability and a credibility score indicating the confidence in the diagnosis.

The estimation of the polyp size influences the diagnosis treatment, so clinicians demands it to be automatically provided by the CAD system, as the polyp size determination through open forceps is subject to error [62]. The challenge is to recover the spatial information for size estimation of a 3D polyp from a colonoscopic 2D image. So far, little progress has been achieved in this regard. A method has been developed for the polyp size classification from colonoscopic videos into under and over 10 mm [63]. Just like the size, the location of the polyp also influences the risk of CRC [64], so it is also interesting to estimate the location of polyps (in mm from the anus), and not only indicate segment they are in.

Besides addressing the above-mentioned clinical and CAD system needs, new imaging technologies with high resolution could also help to reduce the problems of colon lighting for better polyp detection as requested operational/physical needs. CAD systems are mainly focused on the detection of polyps within the endoscopic field of view [54] so their performance is highly dependent on the proper exploration of the colonic mucosa and polyp exposure, as 20% of the colon surface is never surveyed [65]. To solve this limitation, Stanek et al. [66] have developed a CAD system to identify dark or hidden areas of the colon and alert the endoscopist, so that they can be explored more thoroughly. This software assesses the video frame quality and withdrawal spiral motion of the colonoscope to display a green marker in real-time when a quadrant of the image is inspected. Other approach is focused on the development of a 3D map of the colon by predicting the depth from endoscopic images, so the localization of the lumen is used to assist navigation during colonoscopy [67], thus providing a real-time quantitative measure of the colon inspection.

However, besides lighting of the colon, the poor preparation of patients’ bowel prior to colonoscopy highly influences polyp detection. Although new trends of split-dose regimen of bowel preparation has been associated with a better colonic cleansing and adenoma detection, the reticent patient attitude towards this split regimen is still a barrier [68].

To maximize the colon exposure, mechanical add-on devices are also used [54], such as a transparent hood, balloon colonoscopy, or wide-angle colonoscopy [69]. These devices have greatly improved polyp detection rates by expanding the visual field behind the folds. Other accessory device-based systems, such as the Endocuff Vision or EndoRings, manipulate the colon folds using radially extended flexible projections [70, 71]. A cheaper and easy-to-access alternative could be the use of a 3D-printed cap adjustable to standard endoscopes with sideoptics [72], which incorporates two micro-cameras fixed to achieve additional views. Approaches like these could meet the operational/physical limitations of current colonoscopes in terms of flexibility or tip stiffness that makes it difficult to find polyps (especially flat ones) located behind the folds or near the colonic flexures.

Availability of endoscopic systems with better image quality is also requested. Current colonoscopes are mainly based on HD images, but sometimes with this resolution images are still blurred, and it is difficult to distinguish the different tissues. Nowadays, some deep learning algorithms are been implemented for imaging improvement, however they can only process 3–6 frames per second (fps) on 4 K videos and 2fps on 8 K videos, far from real-time requirement (25–30 fps) for endoscopy applications [73].

According to the discussion, different challenges still need to be addressed to meet the identified needs. Regarding clinical needs, the assessment of lesion margins and inspection of the remaining tissue of diminutive polyps with less than 5 mm with imaging techniques should be further improved, with the aim of reducing recurrence rates, since in this kind of polyps it can reach up to 11% [45].

In case of CAD system needs, additional efforts are encouraged in the development of CAD systems for the detection of flat polyps, as they are problematic for endoscopists and have a considerable prevalence of almost 25% [74]. Moreover, a CAD system that estimates the polyp size in real time, rather than classifying it into categories, should be developed.

The first challenge related to operational needs would be the development of a tip with 360° vision, either with a rotating tip or with a system that would allow peripheral as well as focused vision. As far as we know, new scope modalities have achieved a vision of 330°, such as the Full Spectrum Endoscopy colonoscope or the Third-Eye Retroscope and Third-Eye Panoramic devices [75]. Additionally, the implementation of an accelerator algorithm for 4 K and 8 K videos appropriate for real-time detection should be further addressed.

The combination of OCT and MPT photonics technologies is also a promising approach that can offer high sensitivity and specificity for diagnosis [76], representing an unprecedented powerful clinical tool to be used for both CRC early diagnosis and follow-up. These images modalities provide microscopic structural and functional information, which are not provided by NBI, FICE or similar imaging modalities.

Beyond the advances in the development of CAD systems for either polyp detection or classification, it would be desirable to incorporate both detection and classification functionalities into one single system to complement the normal workflow of endoscopists [61]. As preferred by endoscopists, visual cues should be overlaid onto the clinical display as augmented reality in case of detection, and with a “traffic light” display with different colours for the different histological classifications, based on Kudo’s pit pattern. In any case, the final decision on the diagnosis should be made by the clinician, with the CAD software acting as support tool. The opinion of endoscopists indicating that these classification methods (Kudo, Paris, etc.) are not simple and reproducible enough, together with findings from other studies showing that their interobserver variability should be improved [77], make their diagnostic predictions not fully reliable.

On the other hand, educational initiatives for patients to improve their compliance of bowel preparation are encouraged. But due to the foreseeable lack of patient involvement, such educational initiatives should also be addressed to healthcare providers. There are some independent predictive factors of an inadequate bowel preparation, such as diabetes, psychiatric illness, opioid use, active tobacco use, history of inadequate bowel preparation, and Medicaid coverage [78]. Therefore, if healthcare providers or endoscopists previously know the patients with these characteristics, they could foresee it and reinforce their education for a proper bowel preparation.

Regarding operation needs, it has been shown that 4 K ultra HD monitors reduce operative time and intraoperative blood loss in colorectal laparoscopic surgery, so these technologies with higher quality should be widely incorporated in clinical settings for the normal colonoscopy practices, since currently these 4 K ultra HD technologies are mainly available in specialized, high-volume laparoscopic centers [79]. However, the reduction of operative time achieved with these technologies should be taken with caution, as acting quickly with a shorter withdrawal time is also associated with higher interval CRC rates [80].