Facing the challenges of autoimmune pancreatitis diagnosis: The answer from artificial intelligence

Abstract

Current diagnosis of autoimmune pancreatitis (AIP) is challenging and often requires combining multiple dimensions. There is a need to explore new methods for diagnosing AIP. The development of artificial intelligence (AI) is evident, and it is believed to have potential in the clinical diagnosis of AIP. This article aims to list the current diagnostic difficulties of AIP, describe existing AI applications, and suggest directions for future AI usages in AIP diagnosis.

Key Words: Artificial intelligence; Autoimmune pancreatitis; Diagnosis; Machine learning; Deep learning

Core Tip: The current diagnosis of autoimmune pancreatitis (AIP) is challenging and often requires multidisciplinary and comprehensive judgment. Misdiagnosis of AIP in clinical practice is common, so it is necessary to find new diagnostic methods for AIP. The development of artificial intelligence (AI) is evident, and we believe that it holds significant potential as a positive auxiliary tool for AIP diagnosis, with substantial practical value in clinical practice. This article aims to highlight the role of AI in AIP diagnosis, suggesting a promising avenue for future development.

TO THE EDITOR

An article recently published by Gallo et al in the World Journal of Gastroenterology has caught our attention as it systematically and comprehensively discusses autoimmune pancreatitis (AIP), a rare pancreatic disease in the digestive system[1]. Due to its multiple subtypes, complex diagnostic processes, and lack of distinctive clinical features, misdiagnosis happens commonly. The article thoroughly introduces the current definitions, epidemiology, pathogenesis, and multifaceted manifestations of the disease (serology, radiology, endoscopic ultrasound (EUS) manifestations, and histopathologic manifestations). It also presents an overview of the available diagnostic methods and therapies, which can serve as a significant clinical guideline. As the article notes, the most challenging aspect of AIP is its diagnosis. The primary challenges of its diagnosis are mainly reflected in the following aspects.

Serum IgG4 levels are crucial for diagnosing AIP. However, elevated IgG4 levels alone are insufficient for diagnosing AIP, as they may also be observed in other conditions, such as IgG4-related disease (IgG4-RD) involving multiple organ systems[1]. The imaging techniques which are the basis of the diagnosis may also result in misdiagnoses due to AID’s similarities with the features of pancreatic ductal adenocarcinoma (PDAC), and computed tomography (CT) or magnetic resonance imaging (MRI) often show similar findings in these conditions[2,3]. Biopsy, or pathological examination, with negative results does not completely rule out AIP, nor does it provide high accuracy in distinguishing AIP[4]. Therefore, AIP often requires a joint diagnosis, as each feature cannot achieve a final diagnosis on its own. For instance, the Mayo HISORt[5] diagnostic model employs five criteria, combining histology findings, imaging examinations, serological tests, involvement of other organs, and response to steroid therapy. However, this framework also has its drawbacks. For example, when using HISORt, a serum IgG4 level greater than twice the upper limit of normal is considered to have the strongest evidence level; however, although it has a specificity of 99%, its sensitivity is only 53%[6].

AIP can be classified into three subtypes: AIP1, AIP2, and AIP3. The article by Gallo et al[1] provides detailed distinctions involving clinical manifestations, pathological processes, and so on, leading to diagnostic differences between AIP subtypes. AIP2 often shows absent or atypical features such as a lack of elevated serum IgG4 levels. The newly defined AIP3 also presents specific radiological findings, consistently associated with acinar damage and pancreatic volume loss, according to Thomas et al's study[7].

In summary, diagnosing AIP is particularly challenging due to the high specificity of clinical manifestations in each patient. However, misdiagnosis can lead to delaying of interventional treatment. In particular, the distinction between PDAC and AIP is crucial for patient management and prognosis.

For the complex diagnosis of AIP, Gallo et al’s article mentions the search for new diagnostic tools and the reliance on emerging new technologies, such as artificial intelligence (AI), biomarkers (such as serum microRNAs), autoantibodies targeting self-antigens, and advanced imaging techniques[1].

Currently, due to the rapid advances of AI, scientists are actively seeking new methods for AIP diagnosis from AI, which might change the predicament. AI is a mathematical technique that automatically learns and recognizes data characteristics. So far, AI has achieved a series of significant advancements in AIP diagnosis.

Application in CT examinations

Anai et al[8] utilized Support Vector Machine classifier based on CT features, which demonstrated excellent performance (area under the curve = 0.920) in distinguishing focal AIP from PD. Simultaneously, CT-based radiomic features assisted by machine learning (ML) also help differentiate AIP from PDAC. In Park et al's study[9], regions of the pancreas affected by PDAC or AIP and normal tissues were segmented into three-dimensional volumes, yielding 431 radiomic features. Using a random forest algorithm, they achieved a total accuracy of 95.2% in distinguishing AIP from PDAC.

Application in EUS examination

Multiple studies indicate that AI can serve as a valuable aid in endoscopy, leading to more accurate and timely diagnoses and potentially better patient outcomes. Marya et al[10] developed an AI model using convolutional neural network (CNN) on EUS images, achieving a 93% sensitivity and 90% specificity. Compared to traditional diagnostic methods, AI models exhibit higher sensitivity and specificity, effectively distinguishing AIP from PDAC.

Computer-aided diagnosis (CAD) of EUS images, which extracts parameters from EUS images, can also be used to differentiate AIP from chronic pancreatitis (CP). Zhu et al's study introduced a new method, the local ternary pattern algorithm, to enhance the performance of classification models[11]. They selected representative EUS images and extracted parameters from the region of interest, constructing, training, and validating a support vector machine prediction model based on feature combinations. The mean (standard deviation) accuracy was 89.3% ± 2.7%, sensitivity was 84.1% ± 6.4%, and specificity was 92.5% ± 3.3%.

In a recent retrospective study[12], EUS images were used to build a predictive diagnostic model with good discrimination between focal AIP and PDAC, with an area under the receiver operating characteristic curve exceeding 0.95. The model had a sensitivity of 83.7%-91.8% and specificity of 93.3%-95.6% for diagnosing PDAC.

In distinguishing whether pancreatic masses are PDAC, Kuwahara et al's study developed an AI model capable of differentiating pancreatic cancer from non-pancreatic cancers by analyzing EUS images of various types of pancreatic masses using a deep learning (DL) architecture[13]. Addressing data scarcity and correcting the imbalance in disease proportions, the study employed a Deep Convolutional Generative Adversarial Network (DCGAN) for data augmentation. DCGAN is an unsupervised DL model that comprises two CNNs: A generator that produces similar images given random noise vectors, and a discriminator that distinguishes between real and generated images. By inputting disease (CP, AIP, and pancreatic neuroendocrine tumor) images into the DCGAN and training iteratively, the image count for each disease is augmented to 20000. This AI model demonstrates a specificity and accuracy of 0.82 and 0.91, respectively, in diagnosing pancreatic cancer, and a sensitivity of 0.73 in distinguishing AIP.

AI can also identify AIP diagnostic complications through decision trees. Exocrine pancreatic insufficiency (EPI) is a common manifestation of CP and AIP. A study[14] used alternative clinical markers to estimate the presence of EPI in CP or AIP patients. Applying machine learning with decision trees to a retrospective training cohort, the decision tree analysis showed that EPI in CP or AIP patients can be primarily identified based on body mass index (≤ 21.378 kg/m²) and total protein (≤ 7.15 g/dL), aiding in guiding clinical implementation for EPI diagnosis and treatment.

In the future, with advancements in imaging technologies and AI algorithms, the following aspects can be considered for further development in AIP.

Development of an AI-based multidimensional diagnostic model

Based on multidimensional diagnostic perspectives and by integrating AI algorithms with existing frameworks, a more accurate and comprehensive diagnostic model can be developed. In the review of specific diagnostic results, AI can be utilized to enhance accuracy. For instance, the diagnostic accuracy of EUS-guided fine needle aspiration (FNB)/fine needle biopsy (FNB) for AIP is currently not high, at only 68.2%[15].

Studies have already implemented the application of AI in EUS-FNA, where Lin et al[16] developed the AI-ROSE model, which enhances diagnostic rates and accuracy during EUS-FNA. There is currently no AI specifically designed for AIP in EUS-FNA/FNB, but future development could address this disease.

Additionally, more specific biomarkers, such as microRNAs and microbiota, could be incorporated into diagnostic models to make them more comprehensive and improve diagnostic accuracy.

Using AI to identify extra-pancreatic changes in AIP for diagnostic supplements

Patients with AIP often have different clinical symptoms that should be noted for damage to extra-pancreatic organs, which are nonspecific, such as obstructive jaundice and abdominal discomfort[1], with some patients even being asymptomatic. In the face of complex diagnoses, it is advisable to consider auxiliary differentiation from the perspective of complications in the future.

The article by Gallo et al indicates that IgG4-RD AIP1 has a 50% probability of producing extra-pancreatic manifestations, such as IgG4-related sclerosing cholangitis, specifically characterized by bile duct dilation and characteristic thickening of the bile duct wall[17].

Yao et al[18] have developed a DL-based system called BP MASTER, which can provide real-time operational instructions during EUS process, depict the anatomical structure of the bile duct, and estimate the bile duct diameter. It demonstrates an accuracy rate of 93.3%. This suggests that AI can be integrated with biliary structure recognition results to assist in diagnosing biliary changes caused by AIP complications.

Development of AI models combining new examination technologies

With the novel imaging technology, more accurate features can be identified. The 7T MRI and magnetic resonance spectroscopy represent an exciting advancement in magnetic resonance technology[19], offering the potential to enhance spatial, spectral, and contrast resolution. The new technology mentioned in Gallo et al‘s article[1], such as EUS-elastography, can also display more characteristic results, such as increased tissue stiffness.

With greater clarity of imaging, AI can learn and summarize image features more completely and potentially, leading to breakthroughs in AIP classification, differentiation from other diseases, and the discovery of new features.

Expanding data during AI development

Currently, most AI models are facing the problem of insufficient data. The development of AI for AIP diagnosis requires a multi-center, large-scale database. As models develop, typical simulators can be used to generate similar images, addressing the rarity of AIP and the lack of patients.

The application of AI in the clinical workflow of AIP holds significant importance but also faces numerous challenges in practical clinical settings.

The pros and cons of AI in clinical applications

The application of AI in clinical workflows of AIP is of significant importance but faces numerous challenges in practice.

In terms of patient diagnosis and prognosis, AI's powerful identification capabilities allows for rapid recognition of disease traits and provision of diagnostic recommendations when applied to AIP which has relatively high misdiagnosis rates. This enables new dimensions of diagnostic insight, optimization of treatment plans, and the achievement of early intervention. Moreover, it streamlines medical workflows by reducing unnecessary examinations and treatments.

From a standpoint of clinical cost benefits, AI demonstrates long-term economic advantages, despite the high initial investment required for AI technology development. For instance, research indicates that using AI, such as CAD, helps in reducing the incidence and mortality rates of colorectal cancer, ultimately saving overall healthcare costs in the long run[20].

The challenges associated with the practical application of AI in clinical settings should not be underestimated. First, insufficient clinical validation prevails, and most AI models for identifying AIP are still in the experimental phase with a lack of large-scale clinical verification. AI’s diagnostics of AIP require the integration of data from various modalities, such as imaging, laboratory tests, and pathology examinations[1]. Differences in the acquisition timing and spatial resolution of these different modalities can make the fusion and alignment of multi-modal data highly difficult.

Moreover, research on large language models indicates that AI may show weak performance in differential diagnosis. This can reflect inadequacies in the training dataset regarding disease correlation or highlight the limitations of models in replicating the nuanced clinical decision-making process[21].

Implementation in clinical practice

Developing strategies to adapt to clinical workflows is essential for addressing practical challenges posed by AI. The AI system should be optimized first, in order to develop more transparent and interpretable AI systems. This necessitates the involvement of clinical personnel in every stage of medical AI algorithm development, including assessing the algorithm's appropriateness at the experimental design stage and requiring clinical personnel to label data with clinical knowledge during the modeling process. Moreover, some counterintuitive errors may arise during the model establishment and validation phases, necessitating clinical experts' involvement in interpreting and fine-tuning the models.

The development and application of DL-based multimodal data fusion for clinical data have shown promise, with current research focusing on areas such as the integration of medical imaging and electronic health records, as well as structured electronic health records and medical texts, demonstrating the immense potential of DL-based multimodal medical data fusion in the field of medical AI research[22].

Customized solutions for specific diseases should be a future direction for AI applications, which would involve the development of specialized models for particular medical specialties or conditions, refined within specific clinical settings to provide more accurate diagnostic and therapeutic information. Studies have shown that AI tailored to specific medical training outperforms general models in diagnostic accuracy[21]. Taking the AI diagnostic model of AIP as an example, it can provide customized AI solutions based on the examination and diagnostic steps involved in AIP, tailored to the needs and characteristics of different medical institutions, to achieve seamless integration with existing workflows. In summary, AI represents a complex and continuously iterative process.

Facing the current diagnostic and treatment challenges, using AI can enable more precise analysis of AIP characteristics, helping to identify new diagnostic approaches. It is imperative for data scientists and medical experts from diverse specializations to collaborate through the entire process of multimodal AI research and application in medicine. This collaborative effort is essential to ensure that AI-driven diagnostics can truly transition into clinical practice, thereby better serving both medical practitioners and patients. It is believed that with the advancement and application of technology, the clinical diagnosis of AIP will become more scientific and objective in the future.