Application of artificial intelligence in primary gastrointestinal lymphoma: Opportunities, development potential, and future directions

Abstract

Primary gastrointestinal lymphoma (PGIL) is a relatively rare subtype of extranodal lymphoma in clinical practice. Occult primary sites, nonspecific clinical manifestations, and complex imaging and pathological features pose significant challenges to the diagnosis and treatment of the disease. Yang et al reported an in-depth analysis of the clinicopathological characteristics of PGIL from a surgical perspective in World Journal of Gastrointestinal Oncology. It not only provided valuable practical suggestions for the clinical management of this disease but also offered an important opportunity to further explore the diagnostic and therapeutic dilemmas of the disease. With the rapid development and breakthroughs of artificial intelligence technology in medicine, it has offered brand-new ideas for solving the diagnostic difficulties of PGIL, and become a key direction for promoting the precise diagnosis and treatment of this disease. This editorial focuses on discussing the application value, development potential, and research ideas of artificial intelligence technology in the diagnosis and treatment of PGIL, with an aim to provide useful inspiration for clinical researchers.

Key Words: Primary gastrointestinal lymphoma; Artificial intelligence; Multimodal diagnostics; Clinical characteristics; Treatment

Core Tip: As pointed out by Yang et al, primary gastrointestinal lymphoma (PGIL) remains challenging in diagnosis and treatment, with misdiagnosis and suboptimal therapies persisting. Artificial intelligence (AI) shows great potential in optimizing PGIL’s diagnosis (radiomics, endoscopy, and pathology) and treatment (subtype identification, regimen selection, and complication prediction). However, AI faces hurdles like scarce standardized datasets, poor interpretability, and data security risks. With sustained research, cross-disciplinary collaboration, and standardized efforts, AI is expected to become a standard clinical tool to improve PGIL patient outcomes.

This editorial refers to "Clinicopathological characteristics and surgical value of primary gastrointestinal lymphoma" by Yang et al, 2025; https://dx.doi.org/10.4251/wjgo.v17.i11.112089.

---

INTRODUCTION

Primary gastrointestinal lymphoma (PGIL), as a special subtype of lymphoma, accounts for 30%-40% of all extranodal lymphomas and 1%-4% of gastrointestinal malignancies, posing significant challenges to clinicians[1,2]. A recent 25-year retrospective study conducted by Yang et al[3] published in World Journal of Gastrointestinal Oncology systematically analyzed PGIL patients’ epidemiological data, clinical manifestations, imaging features, pathological characteristics, and treatment outcomes. The study identified diffuse large B-cell lymphoma (DLBCL) as the main subtype of PGIL, with abdominal pain being the most common symptom. Diagnosis primarily relies on computed tomography (CT) (with a sensitivity of 94.3%) and endoscopic biopsy (with a detection rate of 91.5%). Surgery plays an important role in lesion resection, complication management, and pathological diagnosis. Beyond these findings, the study provided solid and valuable practical suggestions for the clinical diagnosis and treatment of PGIL, and its results haav significant reference value for optimizing PGIL diagnostic strategies and individualized treatment regimens. However, the study pointed out that even with the combined use of multiple diagnostic methods, including CT and endoscopic biopsy, the problems of diagnostic delay and misdiagnosis remain unresolved. On the other hand, in the field of treatment, the challenge of individualized decision-making further complicated the complexity of management of PGIL. As pointed out in the study, in current clinical practice, surgery is often used as a means for the passive management of tumor-related complications rather than active intervention. The improvement rate of PGIL in patients receiving surgery combined with chemotherapy seemed higher than that in patients receiving chemotherapy alone (40.9% vs 25.0%), but this difference did not reach statistical significance. Additionally, the complexity of tumor and molecular subtypes significantly increases the difficulty of chemotherapy or targeted therapy. However, what is more in line with modern diagnostic and therapeutic concepts is the imperative of precise preoperative assessment of tumor size, invasive extent, and complication risk[4,5]. This assessment is necessary to formulate individualized treatment strategies, including surgical interventions, chemotherapy, or targeted therapies, and confers survival benefits to patients with PGIL ultimately.

Driven by digital technology, artificial intelligence (AI) has evolved exponentially[6,7]. By integrating multi-source data to address tumor heterogeneity and high-dimensional data challenges, AI not only accelerates scientific research progress and innovates oncology research methods but also provides critical technical support for the implementation of precision medicine[8,9]. Therefore, this editorial focuses on the application value, development potential, and research directions of AI in the diagnosis and treatment of PGIL, aiming to provide valuable insights for clinical researchers.

CURRENT DIGNOSTIC CHALLENGES IN PGIL

Given the lack of specificity in the clinical manifestations of PGIL, diagnosing this disease by taking clinical symptoms as the entry point often fails to achieve ideal results. According to the study from Yang et al[3], the diagnosis of PGIL relies on the integration of imaging, endoscopic, and pathological data, which all present opportunities for AI intervention.

AI in imaging

As an emerging technology in the field of medical imaging, radiomics can extract tumor features that are difficult to identify by human vision from images such as CT/magnetic resonance imaging, providing a quantitative basis for the imaging diagnosis of PGIL[10,11]. However, the massive volume and high complexity of radiomic features pose significant challenges to manual analysis. In contrast, machine learning and deep learning algorithms, can automatically and efficiently learn and screen effective features from a vast amount of radiomic data, eliminating redundant information and reducing the interference of subjective factors from radiologists[12,13].

From the perspective of practical application pathways, the implementation of this technology can be advanced through a standardized workflow[14,15]. In detail, deep learning algorithms are leveraged to conduct comparative analyses between PGIL images and imaging data of other common gastrointestinal tumors. Subsequently, differential features are extracted to optimize model training, ultimately leading to the construction of a targeted PGIL diagnostic model[16].

Multiple clinical studies have confirmed that AI models trained on large-scale imaging datasets exhibit significant advantages in lesion detection. They not only enable rapid and accurate identification of suspicious lesions but also effectively capture subtle lesions that are easily overlooked by radiologists[17]. Notably, dedicated radiomics research focusing on PGIL remains relatively scarce at present. Compared with high-incidence gastrointestinal malignancies like gastrointestinal stromal tumors and adenocarcinomas, the accumulation of PGIL-related imaging data and the development of specialized models are significantly lagging behind. This renders AI-integrated radiomics an extremely valuable research direction in the field of PGIL diagnosis, with great significance in filling existing clinical gaps.

However, in practical implementation, a series of issues require attention. First, the incidence of PGIL is relatively low, which necessitates multi-center and long-term data collection. Currently, medical imaging data from different institutions exhibit variations in scanning parameters, image formats, and other aspects, which directly affects the training effectiveness and generalization ability of the models[18]. Therefore, it is essential to establish a unified PGIL imaging database that adheres to standardization protocols. Second, the interpretability of AI models remains a bottleneck[19]. Most deep learning models are classified as “black-box models”, and clinicians struggle to understand the specific reasoning process behind their diagnostic results, which limits the trust and acceptance of this technology in clinical practice[20].

Despite these significant challenges, from the perspectives of technical potential and clinical needs, AI-integrated radiomics is fully equipped with the core conditions to serve as a primary detection modality for PGIL diagnosis. With the continuous evolution of deep learning algorithms, the deepening of multi-center collaboration in data standardization, and the strengthening of interdisciplinary cooperation among fields, the current issues will be gradually resolved.

AI in pathology

As key steps in confirming the diagnosis of PGIL, endoscopy and pathological diagnosis have long been constrained by subjective experience and technical limitations. How to integrate AI technology to break through these bottlenecks and deliver more efficient and accurate diagnostic support for clinical practice is an urgent issue that needs to be addressed at present.

Endoscopy is a crucial means for direct visualization of gastrointestinal mucosal lesions, but its diagnostic efficacy heavily relies on the experience and attention of endoscopists[21]. For PGIL, which often presents with non-specific manifestations such as mucosal erosion, ulceration, or subtle thickening, it is easy to be misdiagnosed as common inflammatory diseases or missed during routine examinations, especially for early-stage small lesions[22]. Notably, AI models trained on large-scale annotated endoscopic datasets can automatically analyze mucosal texture, color, and structural changes in endoscopic images[23,24]. By quickly marking suspicious lesions and issuing early warnings, these models effectively reduce missed diagnosis rates caused by human fatigue or insufficient experience[25,26]. For PGIL, it is essential to further collect typical endoscopic images of PGIL cases (including capsule endoscopy) and establish a dedicated dataset to support model training for future auxiliary detection.

However, AI-assisted endoscopy still faces practical hurdles when applied to PGIL diagnosis. First, endoscopic equipment and examination protocols vary widely across clinical settings, such as differences in endoscope models, light source parameters, and operator techniques. Those all can contribute to significant inconsistencies in image quality and the presentation of lesion features. And these variations directly compromise the generalization ability of AI models. For another, there remains a lack of large-scale, multi-center endoscopic image databases dedicated to PGIL, with standardized annotations that include pathological confirmation results. Thus, future efforts should focus on collecting multi-center, large-sample-sized PGIL endoscopic images with standardized annotations to support the training of robust AI diagnostic models.

Pathological examination is the “gold standard” for PGIL diagnosis and subtype classification, but it is a labor-intensive and time-consuming process. It requires pathologists to manually observe tissue sections under a microscope, identify cell morphology, and judge immunohistochemical results[27]. For PGIL, which has complex pathological subtypes (such as DLBCL, marginal zone lymphoma, and follicular lymphoma), the accurate identification of subtle differences in cell structure and marker expression is extremely challenging[28]. Even experienced pathologists may have inter-observer variability, leading to delayed diagnosis or incorrect subtype classification, which directly affects treatment strategy selection[29].

Encouragingly, by converting traditional pathological sections into high-resolution digital images, AI can quickly scan the entire section and perform quantitative analysis of cell morphology and immunohistochemical marker expression. For PGIL, AI models trained on large-scale pathological datasets can not only automatically identify lymphoma cells and distinguish them from normal lymphocytes or inflammatory cells but also accurately classify subtypes[30]. For example, a study showed that a deep learning model achieved a subtype classification (area under the curve of 0.733-0.856) and accurate differentiation of lung cancer (area under the curve of 0.97), which was significantly higher than the average inter-observer accuracy of pathologists[31].

AI IN TREATMENT DECISION-MAKING

PGIL exhibits significant heterogeneity in pathological subtypes, tumor invasion extent, and patient baseline conditions, rendering individualized treatment decision-making a major challenge in clinical practice[32]. Moreover, conventional treatment strategies rely heavily on physicians’ experience, and surgery is used as an approach to manage complications (e.g., gastrointestinal bleeding and perforation). All above issues may lead to suboptimal treatment regimens. Against this backdrop, AI can serve as a crucial tool to optimize treatment decision-making and strengthen complication management, thereby achieving precise and personalized treatment for PGIL.

Molecular subtype identification and chemotherapy regimen selection

Accurate identification of molecular subtypes and precise selection of chemotherapy regimens are crucial for improving the prognosis of PGIL patients. Deep learning and machine learning can integrate multi-omics data (genomics and transcriptomics) and clinical imaging features to establish predictive models[33,34]. For instance, a recent study used a convolutional neural network to analyze gene expression profiles of PGIL samples, achieving an accuracy approaching 100% in distinguishing DLBCL subtypes, outperforming traditional pathological classification[35]. This non-invasive and efficient approach enables early subtype confirmation. By adopting similar multi-modal data integration strategies (e.g., combining genomic data with radiomic features), future studies can further enhance the accuracy and applicability of subtype identification models. For example, a study integrated five-modal data, including radiological imaging, pathological images, genomics, transcriptomics, and proteomics, and successfully classified gliomas into three subtypes with distinct prognostic characteristics[36].

In PGIL chemotherapy regimen selection, AI can leverage large-scale multi-modal datasets to learn the correlations between molecular subtypes, patient characteristics, and treatment outcomes. The efficacy and toxicity of different regimens for individual patients can be predicted, providing personalized recommendations by deep learning[37]. Similarly, researchers have also reported that machine learning can tailor personalized chemotherapy regimens for tumors through detailed genetic analysis, which holds the potential to enhance treatment efficacy and reduce adverse effects[38,39]. Notably, while current AI models show promise in predictive performance, their clinical utility requires further validation through prospective cohort studies.

Surgical timing selection and complication prediction

For PGIL, surgical timing selection and perioperative complication prediction are intertwined, as improper timing directly elevates complication risks. Traditional approaches, which rely on subjective clinical judgment, often lack evidence-based monitoring mechanisms. This shortcoming makes it difficult to minimize the incidence of surgical complications.

Surgery related complication risks can be identified through machine learning algorithms[38,40]. Appropriate surgical timing can be determined via preoperative clinical interventions, thereby achieving the goal of reducing the probability of complications. Additionally, predictive models can be designed to quantitatively assess the risk of complications requiring emergency surgical intervention by integrating radiomic features and clinical indicators[41]. This approach allows clinicians to distinguish between patients requiring emergency surgery and those eligible for conservative treatment. In the prediction of surgical complications, it is feasible to forecast perioperative risks by leveraging patients’ baseline data and surgery-related data[42,43]. Undoubtedly, this presents a promising research direction.

LIMITATIONS AND FUTURE DIRECTIONS

PGIL remains a diagnostic and therapeutic challenge, but AI offers a path toward precision medicine. As highlighted by Yang et al’s study[3], PGIL’s nonspecific symptoms, complex pathology, and unsatisfied treatment outcome demand innovative solutions. From imaging and pathological diagnosis to optimizing treatment selection and prognostic prediction, AI can improve clinical management of PGIL. However, realizing this potential requires addressing not only established hurdles but also emerging concerns surrounding AI monitoring. First, there is a lack of dedicated datasets for PGIL, including pathological images and radiological images. In addition, due to the low incidence of PGIL, the initially trained AI models are at risk of bias, which requires multi-center, large-sample datasets for repeated training and correction. However, the application and popularization of the model across multiple centers are confronted with practical challenges. Second, continuous AI training relies on accessing sensitive patient information, raising risks of data breaches or unauthorized use if robust security protocols are lacking. Third, the “black-box” nature of AI might hinder clinicians’ decision-making, potentially leading to either ignored critical signals or unnecessary interventions[44,45]. Additionally, over-reliance on AI monitoring could undermine clinicians’ intuitive judgment. This risk is heightened in complex PGIL cases, where nuanced clinical context may not be fully captured by algorithms.

Given the conservatism and safety priorities inherent in clinical practice, over-reliance on AI for making critical judgments is decidedly inadvisable. Thus, the future direction of AI in PGIL is not to replace clinicians, but to serve as a powerful auxiliary tool that supports their professional judgments. However, based on the current research foundation, a great deal of work still needs to be further accomplished in the future. The application of AI in healthcare appears to be irreversible. Only through sustained investment in extensive research and the implementation of rigorous clinician training can we ultimately achieve the goals of improving diagnostic accuracy, reducing the incidence of complications, and enhancing patient outcomes.

CONCLUSION

In conclusion, AI demonstrates great potential in the diagnosis and treatment of PGIL. However, AI also faces challenges such as a lack of standardized datasets, insufficient interpretability, and data security risks. In the future, it is hoped that through further research, interdisciplinary collaboration, and standardization construction, AI can assist clinicians in further improving the prognosis of patients with PGIL.