JMIR Publications

ABSTRACT

Background:

Lymphoma is a malignant tumor originating from the lymphoid hematopoietic system and is one of the most common hematological tumors worldwide. According to epidemiological data, Hodgkin's lymphoma (HL) and non-Hodgkin's lymphoma (NHL) are common malignant lymphatic diseases that threaten public health. GLOBOCAN 2020 statistics show that the estimated number of HL cases and deaths worldwide is 83087 and 23376, respectively, while that of NHL is 54 352 and 25 993, respectively. Histopathology, the examination of tissue specimens at the cellular level, is the gold standard for lymphoma diagnosis. The conventional diagnostic process is that pathologists usually use hematoxylin-eosin (H&E) stained tissue to make a diagnosis, but the diagnosis has the disadvantages of subjectivity and time-consuming. In diagnosing difficult cases, the general pathologist may seek the help of a subspecialty lymphoma pathologist, and/or use ancillary tests such as immunohistochemistry (IHC). Referral and ancillary testing are critical to the accuracy of the diagnostic process, but at the cost of making its diagnostic cycle longer and more expensive. It is becoming increasingly common for pathologists to use computers to review whole slide images (WSI) of analytical scans. These tools can often improve diagnostic accuracy, efficiency, objectivity, and consistency. These tools can help alleviate the global pathologist workforce shortage, improve diagnostic throughput, and reduce the need for referrals and ancillary testing

Objective:

Artificial intelligence (AI) has great potential in the diagnosis, prognosis, or gene prediction of lymphoma. We aimed to summarize the performance of AI models for diagnostic or prognostic purposes based on published studies on histopathological images used in lymphoma.

Methods:

This study followed the Systematic Review Reporting Program guidelines. A literature search of PubMed, Cochrane Library, and Web of Science was conducted from inception until February 28, 2024. Included in the standard requirement will artificial intelligence for the prognosis of human lymphoma tissue pathology images or diagnosis, gene mutation, etc. The risk of bias was assessed using PROBAST. Information for each model was tabulated, and summary statistics were reported. The study was registered with PROSPERO (CRD42024537394) and followed PRISMA 2020 reporting guidelines.

Results:

The search identified 3414 records, of which 31 articles were eligible for inclusion. With a total of 42 models, these studies included 16 diagnostic models, 8 prognostic models, 1 model to detect whether the gene was ectopic, and 17 other diagnostic related models. Common tasks include identification task (6/42), classification task (31/42), segmentation tasks (5/42). Models were developed using 10 to 84139 histopathology slides from 10 to 1005 lymphoma patients. In all studies have found higher risk bias or not clear. In the high-risk model, high-risk scores appeared in the participant, outcome, and analysis sections (3/31). Most of the low-risk domains appeared in the predictors (21/31) and outcomes (18/31) sections. Almost all papers had an unclear risk of bias in at least one domain, the most common being the domains of participants (15/31) and statistical analysis (29/31). In the overall practicality evaluation, most of the models were unclear (16/31), high-risk models (12/31), and low-risk models (2/31). the most common reason is the analysis of the participant recruitment limited and no interpretability of the results analysis.

Conclusions:

applying artificial intelligence to lymphoma tissue pathology diagnosis or prognosis of the purpose of image limited study, found no one model can prove to be prepared for the implementation of the real world. Key aspects of accelerating the clinical translation of AI include comprehensive reporting of data sources and modeling approaches, interpretability of AI models, and improved quantitative assessments using cross validation and external validation. Clinical Trial: The study was registered with PROSPERO (CRD42024537394)