Computer-aided diagnosis: further eliminating the human factor in endoscopy?

Advanced imaging modalities in endoscopy provide visualization of the tissue in greater detail than what can be obtained by gross visualization through white-light endoscopy. Magnifying chromoendoscopy, digital chromoendoscopy with magnification, confocal laser endomicroscopy, and ultra-high endocytoscopy provide images that enhance mucosal, vascular, or even cellular visualization. Advanced imaging technologies provide an image, but that image is only useful if it can be correctly interpreted in a clinically meaningful way.

In the colon, advanced imaging has the potential to improve a range of clinical outcomes. For example, the use of narrow-band imaging may be used to diagnose diminutive colon polyps, and meets thresholds to consider a resect and discard strategy to minimize cost, time, and risk [1]. On the other end of the spectrum, advanced imaging with endocytoscopy may provide information to determine which lesions are amenable to endoscopic resection by providing a real-time method to diagnose invasive cancer [2].

“We need to define what standards, if any, are required, and which disclaimers are appropriate in adopting computer-generated diagnostic strategies.”

The operator-dependent nature of the image interpretation is a challenge in the advanced imaging arena. Development of standardized criteria and simplification of classification systems allow for improved communication and applicability. Validation of these systems is required, and the process may be iterative and need refinement along the way. Learning curve, competency, confidence level, and time to diagnosis may lead to variability and obstacles at the endoscopist level. Furthermore, the performance trials are often conducted with expert operators in referral centers, thus limiting the generalizability of the performance to the broader community. To further minimize the operator dependence of image interpretation, computer-assisted algorithms or computer-assisted diagnosis (CAD) have been proposed to provide a real-time diagnosis.

Takeda et al. present their work on a CAD system for ultra-high magnification endocytoscopy (EC-CAD) [3]. They used the first data set of 5543 images for machine learning. Then they tested the algorithm on a set of 200 images for the diagnosis of invasive cancer. The EC-CAD system analyzed the images based on nuclei and texture analysis. A diagnosis of non-neoplasm, adenoma, and invasive cancer was achieved within 0.3 seconds if the image was analyzable. They achieved a sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of 89.4 %, 98.9 %, 94.1 %, 98.8 %, and 90.1 %. These parameters were higher in the setting of a high-confidence diagnosis. The authors conclude that an accurate determination of invasive cancer with EC-CAD will enable the endoscopist to appropriately choose between endoscopic resection or surgery, and thereby reduce unnecessary complications encountered during endoscopic resection of high-risk lesions.

The CAD approach provides a vehicle for generalizability of advanced imaging modalities to be incorporated among a broader population of endoscopists. First, the technical proficiency of complex endoscopic procedures does not automatically parallel the cognitive competency for interpretation of histology-like images. Second, even if the endoscopist may be trained to interpret images, there is still variability associated with competency, time to interpret, and confidence of interpretation. Efficiency, reproducibility, and support for documentation may be better achieved with CAD.

Computer algorithms can also be designed to take advantage of processes that we rely on as humans. Most of us are familiar with digital measurements of spatial parameters, such as diameter and area, but computer software algorithms may now also differentiate recognizable patterns with texture analysis, mimicking aspects of how humans may use pattern recognition for diagnosis. Furthermore, the more complex the classification system is, the more cumbersome, and it is more difficult to implement across a range of users. A computer is designed to handle the complexity provided the complexity is incorporated into the algorithm.

Given the rapid advancement of imaging technologies and the potential of computer software algorithms, how far are we from adopting CAD to eliminate the human factor in image interpretation? At the rate of the evolution of CAD in studies such as the one by Takeda et al., it is not hard to imagine the incorporation of CAD in a variety of imaging modalities that provide an optical biopsy. As users and as a medical community, it is helpful to consider points as we incorporate computer algorithms into daily diagnostics.

First, it is important to acknowledge that computers and humans have different frames in which they operate. The computer is limited by what was provided during the algorithm development, and this selection during teaching will limit the “frame” of the computer’s “understanding.” The algorithm is then limited by how “artificial” the teaching set may be. A teaching set with high-quality images from classic pathology without “noise” or artifact will not capture what we encounter in the real world. Endoscopists may incorporate into their frame of a working diagnosis the understanding of disease processes, differential diagnosis, and appreciation of classic vs. atypical presentations. Seasoned endoscopists with excellent development of pretest probability, given the clinical scenario and the endoscopic findings combined with true knowledge of the imaging platform, competence in interpretation, and appreciation of the limitations, will likely have an accurate gestalt. However, not everyone is an expert, and CAD allows image interpretation to extend beyond the experts to nonexperts and trainees. Consideration and investment need to be made into developing an appropriate framework in order for the computer models to be reflective of the real world.

Second, most algorithms will plateau in performance as dictated by the instructions of the algorithms, whereas most humans learn from experience and improve in performance after multiple encounters. The trajectory of the human learning curve is likely steeper than a set computer algorithm, which is likely to be a steady execution from the algorithm development. Furthermore, there are some confounding effects that are common not only in endoscopy but also even in histology, such as architectural distortion from inflammation. If pathologists have trouble with some of these confounding variables, we should question how a computer algorithm may overcome those variables. CAD may potentially overcome some of these challenges if we go beyond a teaching set that informs a set algorithm to the broader concept of machine learning where the computer is iterating rapidly and continually such that a plateau does not occur.

Third, we need to define the outcome that we are proposing to answer with CAD and be mindful of this as the user. We may construct an ideal question along a binary or linear continuum, such as non-neoplastic vs. neoplastic or a continuum of non-neoplastic – adenoma – invasive cancer. But that does not reflect the whole array of very different diagnostic pathologies that may be encountered. For example, sessile serrated polyps may or may not be incorporated easily or discretely into these continuums, but these lesions are frequently seen in this clinical context. Also, we may encounter other pathologies outside of the context of a specific disease process. This varies organ to organ and among models of pathogenesis. The pathology encountered may be a myriad of disease processes, and image patterns may not fit neatly into discrete groups all of the time.

Finally, as a community we will eventually need to consider accountability and medicolegal implications of CAD. We need to define what standards, if any, are required, and which disclaimers are appropriate in adopting CAD strategies as a part of our clinical practice.

The emergence of the optical biopsy during our endoscopic procedure is now a reality. CAD has the potential to broaden the application of advanced imaging modalities beyond the experts at a few centers to a range of users in the community. As we embrace CAD, we should be mindful of the limitations, and be consistent in using it in the context in which it has been developed.

• References

• 1 Abu DayyehBK, Thosani N, Konda V. et al. ASGE Technology Committee systematic review and meta-analysis assessing the ASGE PIVI thresholds for adopting real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc 2015; 81: 502
• 2 Kudo SE, Wakamura K, Ikehara N. et al. Diagnosis of colorectal lesions with a novel endocytoscopic classification – a pilot study. Endoscopy 2011; 43: 869-875
• 3 Takeda K, Kudo S, Mori Y. et al. Accuracy of diagnosing invasive colorectal cancer using computer-aided endocytoscopy. Endoscopy 2017; 49: 798-802