Optimizing endoscopic ultrasound guided fine needle aspiration through artificial intelligence.

Rapid on-site evaluation (ROSE) improves the diagnostic yield of endoscopic ultrasound-guided fine needle aspiration (EUS-FNA). The cytopathologist performs direct visualization of the cytological smear by microscopy to confirm specimen adequacy and provides a preliminary diagnosis. Further sampling is performed immediately if the specimen is inadequate, thus avoiding the need for repeat procedures.1 ROSE can also reduce the number of unnecessary needle punctures during EUS-FNA, as the procedure is stopped once adequate tissue sample is obtained, potentially reducing the risk of local complications. ROSE is not widely available in many centres in Asia, due to logistical challenges of having a cytopathologist in attendance during all EUS-FNA procedures, and it also incurs additional manpower cost. In recent years, there have been significant advances in artificial intelligence (AI)-based systems, which use algorithms to perform tasks that would usually require human intelligence and input, and AI has now been incorporated into medical practice.2 AI-assisted colonoscopy is now part of regular clinical practice, and it has been shown to significantly improve adenoma detection rate.3 There is ongoing research into the use of AI in upper endoscopy, EUS and capsule endoscopy for lesion detection and characterization, as well as quality tracking.4 The introduction of digitalized whole slide image technology has enabled the application of digital pathology in clinical practice and education. This digital transformation has in turn facilitated the implementation of AI-based algorithms in the field of histopathology in the context of AI-driven cancer diagnostic, prognostic, and predictive applications.5 Published literature on the application of AI-models in the field of cytopathology are fewer in comparison, with a focus on gynecological samples initially. In a systematic review on recent application of AI in non-gynecological cancer cytopathology, only 28 published original studies with full texts from 2010 to 2021 were identified. These studies involved 8 target organs and only one study involved the pancreas.6 It is a natural progression to explore application of AI support systems during the EUS-FNA procedure. This represents the interface between endoscopy and cytopathology, in the context of real-time assessment of adequacy of tissue sample and providing a provisional diagnosis before formal histopathological assessment. In this issue of Journal of Gastroenterology and Hepatology, Lin et al. reported the validation of a novel AI-based model (ROSE-AI model) to substitute for ROSE during EUS-FNA.7 Digitized images from Diff-Quik-stained EUS-FNA slides from 51 patients, predominantly for pancreatic lesions (94.1%), were used to train and validate the AI-algorithm, with an emphasis on specificity for malignancy. The ROSE-AI model achieved an accuracy of 83.4% in the internal validation dataset and 88.7% in the external test dataset. The sensitivity and positive predictive value were 79.1% and 71.7% in internal validation dataset and 78.0% and 60.7% in external test dataset, respectively. The specificity and negative predictive value were 85.4% and 89.7% in the internal validation dataset and 90.6% and 95.7% in the external test dataset, respectively. The accuracy for the training dataset of this AI-ROSE model is 97.7%, which is comparable to the model described in an earlier non-validated study involving 75 FNA pancreas samples that highlights its potential in improving the diagnosis for cases that fall within the atypical diagnostic category.8 The study by Lin et al. is one of the few to evaluate the application of AI-model in diagnostic cytopathology in the context of samples obtained from EUS-FNA of pancreatic lesions, and the first to demonstrate the potential feasibility of using an AI-system to replace manual ROSE during EUS-FNA. The study design was appropriate with the appropriate method of algorithm development in terms of training, internal and then external validation. However, with a sensitivity performance of under 80% in both the internal and external validation datasets, the results are still suboptimal, and this study remains a proof of concept. Additional issues also need to be addressed before it can be translated into clinical practice. When EUS-FNA is performed, the key issues are firstly whether adequate lesional tissue sample has been obtained, and secondly to establish if the acquired sample is truly benign or malignant. The sole focus of the current algorithm is on detection of malignant cytology. To replicate ROSE more closely, the ideal AI-model should be able to assess adequacy of sample acquired especially in the setting of blood contamination and provide a ‘benign’ or ‘malignant’ diagnosis with high confidence. In the current model, there is the need for high-quality staining, slide scanning and digitalization before running the machine-learning algorithm. These requirements present additional practical and financial obstacles for infrastructure development within the endoscopy unit and/or pathology laboratory. Digital pathology also has inherent disadvantages of additional workflows, need for additional personnel, and need for data storage space, which further increase the operational cost and procedure time. During ROSE, the experienced cytopathologist performs microscopic examination on freshly prepared slides without coverslips when they are often still wet to achieve adequacy check and provide a preliminary diagnosis usually under 5 minutes. In their discussion, the authors made a general statement about the AI-ROSE model improving EUS-FNA efficiency without taking into consideration the additional time required for drying (slides must be fully dried before loading into the scanner to avoid residue getting onto the scanner mechanism) and cover-slipping the Diff-Quik-stained slides, slide scanning and digitalization prior to the process of running the machine-learning algorithm. This extra operational time may not be practical for the endoscopist and may potentially affect the safety of the patient who is under sedation. The significant cost in setting up a digital pathology system has limited its widespread adoption beyond well-funded academic medical centres where access to ROSE is likely to be available already. The ideal AI-ROSE support system should require less set-up costs and infrastructure support and should not significantly increase the procedure time. One possibility may well be building such AI-algorithms into portable devices, such as smartphones, that can simply make prediction on sample adequacy and presence of malignant cells by capturing and examining images from Diff-Quik-stained EUS-FNA slides. Regarding challenges about medicolegal responsibility and regulatory issues of an interim report by an AI-ROSE model, another more practical solution is engaging an off-site cytopathologist by way of telepathology with the use of digital pathology. Current alternative strategies to ensure specimen adequacy include macroscopic on-site quality evaluation (MOSE) by the endoscopist to ensure at least 4 mm of macroscopic visible core tissue,9 and indirect methods without tissue visualization, such as performing a pre-determined minimum number of needle passes, which has been correlated with obtaining adequate tissue sample,10 and the use of biopsy needles, which can increase the amount of core tissue.11 Although MOSE is useful, it does not directly assess whether the lesional tissue was sampled, and not adjacent normal tissue, and it just provides an indication of adequacy of tissue sample for further analysis. This holds true for the two indirect approaches in that they are surrogates to ensure adequate tissue has been obtained and do not differentiate between lesional or non-lesional tissue and cannot provide a provisional diagnosis real-time. However, all these approaches have served us well in actual clinical practice and even as we seek to explore new solutions, we can still rely on them in our practice. Cytology smears alone may be insufficient for complete diagnosis. Additional material may be required for ancillary immunohistochemistry and molecular studies for diagnosis and personalized therapeutics. Nonetheless, development of the AI-ROSE model is a positive step forward and when refined and available for clinical practice, it will further expand our armamentarium to improve the diagnostic yield of EUS-guided tissue acquisition.