Lymph node disease in 2-deoxy-2-fluorodeoxyglucose positron emission tomography/computed tomography imaging: Advances in artificial intelligence-driven automatic segmentation and precise diagnosis

Abstract

Imaging evaluation of lymph node metastasis and infiltration faces problems such as low artificial outline efficiency and insufficient consistency. Deep learning technology based on convolutional neural networks has greatly improved the technical effect of radiomics in lymph node pathological characteristics analysis and efficacy monitoring through automatic lymph node detection, precise segmentation and three-dimensional reconstruction algorithms. This review focuses on the automatic lymph node segmentation model, treatment response prediction algorithm and benign and malignant differential diagnosis system for multimodal imaging, in order to provide a basis for further research on artificial intelligence to assist lymph node disease management and clinical decision-making, and provide a reference for promoting the construction of a system for accurate diagnosis, personalized treatment and prognostic evaluation of lymph node-related diseases.

Key Words: Lymph node metastasis; Lymphoma; Deep learning; Convolutional neural network; Medical imaging analysis; Automatic segmentation; Radiomics

Core Tip: This paper reviews the progress of 2-deoxy-2-fluorodeoxyglucose positron emission tomography/computed tomography lymph node disease diagnosis technology driven by artificial intelligence. The automatic segmentation technology based on deep learning has significantly improved the diagnostic efficiency and consistency in lymph node detection, precise segmentation and three-dimensional reconstruction, and made up for the shortcomings of poor efficiency and obvious subjectivity in traditional artificial segmentation. The deep learning model has performed well in predicting treatment responses, distinguishing benign and malignant lesions, and diagnosing lymph node metastasis in various cancer types, providing technical support for the accurate diagnosis of lymph node diseases, individualized treatment and prognostic evaluation.

INTRODUCTION

Cancer is considered one of the most difficult public health challenges in the world. Its mortality rate is only lower than that of cardiovascular disease, ranking second among the causes of death globally[1]. As a key component of the immune system, lymph nodes are mainly responsible for the filtration of lymph fluid and identification of pathogens. Research in recent years has pointed out that lymph nodes play an important role in tumor metastasis and patient prognosis, especially in malignant tumors such as lymphomas that originate from the lymphatic system. Abnormal lymph node function is closely related to the occurrence of diseases. One of the main pathological features of cancer is lymphatic infiltration, which is usually accompanied by lymph node metastasis. The occurrence of lymph node metastasis directly affects the survival rate of patients and greatly increases the risk of cancer-related death. Therefore, the diagnosis and staging evaluation of lymph node metastasis has become one of the decisive factors in evaluating the prognosis of cancer patients and has a key reference value for clinical decision-making[2,3].

As an advanced imaging technology that combines function and anatomical information, positron emission tomography/computed tomography (PET/CT) has been widely used in the diagnosis and staging of a variety of malignant tumors. Deep learning[4] is a machine learning method based on multi-layer neural networks, among which convolutional neural networks (CNNs)[5] are the most frequently applied deep learning network model in the clinical field. Traditional imaging diagnosis such as ultrasound (US), CT and magnetic resonance imaging (MRI) often rely on morphological standards and have limited sensitivity to early metastasis[6]. By evaluating lymph node metabolic activity, we can find out the metabolically active lesions that may be missed by conventional imaging techniques. Several studies have shown that the sensitivity of 2-deoxy-2-fluorodeoxyglucose (¹⁸F-FDG) PET/CT for detecting metabolically active lesions is usually higher than that of US, CT or MRI alone[7-10]. In addition, traditional imaging diagnosis also has pain points such as inefficiency, strong subjectivity, and difficulty in processing big data, and artificial intelligence (AI) technology can effectively solve these problems. In terms of lesion detection and segmentation, AI can realize fully automatic identification or assist doctors in semi-automatic segmentation[11,12]. Deep neural networks show excellent capabilities when processing large amounts of scanning data, which can maintain the accuracy and reliability of the diagnosis, or improve it, and achieve results faster[13]. Radiomics is a rapidly developing research direction. Its core is to extract quantitative features from medical images and then transform this information into complex high-dimensional data[14]. The learning process of AI often starts with data collection and preprocessing, and then feature extraction, model establishment and effect verification are carried out in turn. After in-depth analysis of these high-dimensional data, it can not only reveal unique biological laws, but also deepen our understanding of the pathogenesis of the disease and support clinical decision-making[15,16].

This review focuses on the latest advances in the use of AI and deep learning methods in ¹⁸F-FDG PET/CT imaging for lymph node-related diseases. This review involves four main categories: (1) Deep learning algorithms containing CNNs and advanced neural network architectures to achieve automatic segmentation of lymph nodes; (2) Differentiation and classification processing of different cancer types, including solid tumors such as lung cancer, esophageal squamous cell carcinoma (ESCC), head and neck cancer, and prostate cancer; (3) Distinction and identification of lymphoma, especially the subtypes of Hodgkin's lymphoma (HL) and non-HL (NHL), involving the identification of diffuse large B-cell lymphoma and other types of diseases; and (4) Evaluation of therapeutic effects and prognosis, involving response evaluation of conventional chemotherapy, chimeric antigen receptor T (CAR-T) cell therapy and other emerging therapeutic methods. These studies show the clinical utility and technological development of AI-driven approaches in enhancing diagnostic accuracy, workflow efficiency, and promoting patient prognosis in lymph node disease management.

RESEARCH ON AUTOMATIC LYMPH NODE SEGMENTATION TECHNOLOGY BASED ON DEEP LEARNING

The diagnosis and treatment of solid tumor lymph node metastasis and lymphoma depend on the accurate determination of systemic disease burden. Lesion segmentation is an extremely critical step. Segmentation[17] mainly aims to accurately locate and depict the lesion boundaries in medical images such as ¹⁸F-FDG PET/CT, which is extremely critical for the clinical diagnosis of lymph node lesions and the formulation of treatment plans. Relying on accurately capturing lesion information from imaging data such as ¹⁸F-FDG PET/CT can provide clinicians with reliable decision-making evidence, which is conducive to distinguishing whether lymph node lesions are benign or malignant, and then implementing personalized treatment to promote the patient's treatment effect and prognosis level. Based on segmentation results, a series of quantitative indicators can be calculated, including metabolic tumor volume (MTV), total MTV (TMTV), total lesion glycolysis (TLG)[18], mean standardized uptake value (SUVmean), partial volume corrected metabolic volume product, whole-body metabolic burden[19], metabolic heterogeneity[20] and lesion dissemination (Dmax)[21,22]. These indicators are of significant value in predicting the prognosis of patients with lymph node lesions and evaluating treatment effects[23-25].

Personalized treatment of lymph node lesions guided by PET imaging relies on reliable and precise lymph node lesion segmentation technology to achieve comprehensive quantification of the disease. Traditional lymph node lesion segmentation is mainly drawn by doctors, and is easily disturbed by subjective factors, resulting in significant differences in segmentation results between different operators. The distribution of lymph tissue in the human body covers a wide range, which further limits the efficiency and accuracy of manual segmentation. With the rapid development of deep learning technology, automatic detection and segmentation technology based on deep learning has gradually become a hot topic in research related to lymph node lesion analysis with its own efficiency and consistency. It is expected to break through the constraints of traditional methods and further improve the diagnosis and treatment level of lymph node lesions. This section will systematically examine new progress in the research of the automatic segmentation technology of lymph node lesions based on deep learning from the aspects of semi-automatic segmentation, fully automatic segmentation, segmentation method comparison and influencing factor analysis.

Semi-automatic segmentation

At present, the semi-automatic segmentation methods are mainly divided into two categories: One is to manually identify and mark the lesion area, and then use the algorithm to automatically segment the marked area. Zhu et al[26] proposed a semi-automatic lymphoma segmentation method based on deep learning and designed a new cross-structure-guided boundary optimization network. The method is characterized by extracting the cross-shaped structural information of lymphoma sections from PET images as additional input, while combining the boundary gradient loss function to improve the accuracy of tumor segmentation boundaries. On lymphoma datasets and publicly available head and neck datasets, the performance of this method is superior to other advanced semi-automatic segmentation methods and is able to generate high-quality and reliable segmentation results.

Another type is that the algorithm first automatically detects and divides the suspected lesion area, and then uses manual review to confirm and eliminate false positive results. Yu et al[27] proposed a semi-automatic detection and segmentation method of lymphoma based on a fully connected conditional random field model, which combines metabolic and anatomical information provided by PET/CT images. Specifically, this method first uses multi-map segmentation technology to eliminate physiological high metabolic organs in CT images, thereby reducing the false positive rate; then, a fully connected conditional random field model is used to perform automated detection and segmentation of lymphoma in the area where high metabolic organs are removed. This method achieved a lymphoma detection rate of 100% in 11 patients, and the mean Dice similarity coefficient in the segmentation results reached 84.4% compared with manual annotation.

Based on the previous analysis, although the semi-automatic segmentation method has outstanding advantages in segmentation accuracy and operational efficiency, its own characteristics determine that there are certain limitations. This method requires manual interactive intervention, which will definitely affect the efficiency of large-scale data processing; subjective differences between different operators may cause inconsistent results. It is worth noting that in the study of metastatic foci segmentation of lymph node metastasis in malignant tumors, fully automatic segmentation methods dominate, whereas semi-automatic methods are relatively rare. This phenomenon may be due to the excessive operational process of semi-automatic segmentation in dealing with a large number of lymph node metastasis lesions, which is difficult to meet the efficiency requirements of clinical practice.

Fully automatic segmentation

With the rapid advancement of AI technology, fully automatic AI models have shown significant advantages in the field of medical imaging analysis. This type of model can effectively enhance the efficiency of clinical workflows and allow nuclear radiologists to free themselves from the complex lesion recognition tasks. Given the wide distribution area and large number of lymphoma lesions, traditional manual recognition methods are not only time-consuming, but also difficult to effectively carry out in routine workflows.

In order to test the practical application value of AI in PET image analysis, researchers have implemented a series of targeted studies, focusing on the comparison between automatic segmentation systems and traditional manual methods. For instance, Pinochet et al[28] conducted a study on a CNN-based PET-assisted reporting system, focusing on its application value in the identification of suspected cancer lesions and in TMTV evaluation. The study analyzed the data of 119 patients, and the results showed that the Dice similarity coefficient between automatic segmentation and manual segmentation was 0.65, and the intra-class correlation coefficient between the automated TMTV and the artificial method was 0.68. It is worth noting that the TMTV obtained by these two segmentation strategies can effectively predict the patients' progression-free survival (PFS) and overall survival (OS). These results show that there is some correlation between automated TMTV and manual segmentation results. In the quantitative analysis of prostate-specific membrane antigen (PSMA) PET/CT images, the application of automated analysis software can enhance the standardization and consistency of the analysis. Johnsson et al[29] analyzed and verified the aPROMISE software, which can perform organ segmentation on low-dose CT images with deep learning algorithms, and can also detect and quantify potential pathological lesions in PSMA PET/CT. When the work of bone segmentation is applied, the Dice coefficient enters the range of 0.88 to 0.95; for reference organs related to PSMA PET/CT, such as the thoracic aorta and the liver, the Dice coefficients are valued at 0.89 and 0.97, respectively. In terms of lesion detection, the sensitivity of the software to detect regional lymph nodes, all lymph nodes and bone metastasis was 91.5%, 90.6% and 86.7% respectively. These results confirm that when aPROMISE software conducts PSMA PET/CT image analysis, it shows good segmentation accuracy, consistency of quantitative evaluation and high sensitivity to lesion detection compared with traditional manual segmentation.

As the research progresses step by step, researchers began to focus on multicenter studies with larger sample sizes to verify model reliability. Given this realistic background, Blanc-Durand et al[30] extracted the complete baseline ¹⁸F-FDG PET/CT image data of 733 diffuse large B-cell lymphoma (DLBCL) patients from two multicenter prospective clinical trials, and developed a fully automatic segmentation model based on three-dimensional (3D-CNN). The average Dice similarity coefficient and Jaccard coefficient of the verification group were 0.73 ± 0.20 and 0.68 ± 0.21, respectively. In the verification process of two independent cohorts, even though there was a certain degree of underestimation, the average underestimation of TMTV in the first cohort was 12 mL (2.8%) ± 263 (P = 0.27), and the average underestimation of TMTV in the second cohort was 116 mL (20.8%) ± 425. However, this model can still show good performance in automatically clarifying lymphoma lesions, especially in TMTV prediction. In another study[31], AI methods based on PET and CT imaging also showed important value in the diagnosis and prognostic evaluation of head and neck tumors. The researchers developed a comprehensive framework to analyze large-scale data sets (524 cases in the training set and 359 cases in the test set) from 9 centers, and achieved automatic segmentation of primary tumors (GTVp) and lymph nodes (GTVn). This approach obtained Dice scores of 0.774 and 0.760 on the test set, respectively. It is worth noting that the direct introduction of uncertainty estimation mechanism can greatly reduce the false positive rate: GTVp has a 19.5% reduction, and GTVn has a 7.14% reduction. The radiomics characteristics of GTVn in PET scan and the characteristics of GTVp and GTVn under CT scan have key predictive value for the prediction. By integrating these radiomics characteristics with clinical characteristics and combining SHAP analysis methods, a prediction model with a c-index of 0.672 can be established, which can be used to stratify patients by risk.

In order to further enhance the value of AI models in clinical practical applications, researchers have begun to look for a comprehensive solution to integrate tumor segmentation and prognosis prediction. A study using deep learning technology[32] was conducted on DLBCL patients in three-dimensional ¹⁸F-FDG PET/CT images, so as to achieve accurate segmentation of tumor areas and predict TMTV. The fully automatic tumor segmentation technology based on deep learning and the TMTV prediction tool showed high reliability and stability in the training set and verification set. After further analysis, it was found that the prediction results generated by the deep learning model were significantly correlated with the patients' PFS and OS, which not only reflected the patient's tumor burden status, but also provided important reference for prognostic evaluation. In the risk stratification analysis, TMTV obtained based on model prediction can effectively divide patient groups at different risk levels. This stratification method is closely related to the patient's survival outcome, confirming the clinical application significance of TMTV based on model prediction values in prognostic prediction. The application of deep learning technology in three-dimensional ¹⁸F-FDG PET imaging analysis not only improves the accuracy and efficiency of tumor segmentation, but also provides reliable quantitative content for the prognostic evaluation of DLBCL patients.

The fully automatic segmentation model relying on deep learning achieves high efficiency and accurate tumor segmentation, and also provides reliable predictive relevant indicators. However, considering that fully automatic segmentation often requires large samples to carry out model training, the fully automatic segmentation model may perform poorly in handling rare cases. Moreover, under different equipment or collection parameters, the generalization ability is limited. In the future, when conducting research, adaptive learning algorithms can be focused on optimizing the generalization level of the model, and at the same time, the method of using small sample training models to deal with the problem of rare case segmentation is explored.

Segmentation comparison

Image segmentation is a key step in the quantitative analysis of disease, and its accuracy directly affects subsequent feature extraction and analysis results. However, there is currently a relative lack of systematic research on the effects of different segmentation methods on lesion feature extraction[33]. When evaluating the reliability of automated lymphoma detection and segmentation models, the manual segmentation results from experts are often used as the gold standard. However, since the lesion status of lymph nodes is difficult to clearly define and the final diagnosis relies on invasive methods such as biopsy or autopsy, it is limited to assessing the performance of deep learning models only through true positivity rates. More importantly, it is necessary to evaluate the model's ability to identify expert consensus areas, that is, the model's identification accuracy of lymph node areas commonly agreed upon by experts, so as to better reflect the reliability and clinical application value of the model.

Previous studies have shown that the performance differences among segmentation methods are obvious. In order to deeply explore the differences in the performance of different segmentation methods in ¹⁸F-FDG PET/CT image processing, the researchers conducted a series of comprehensive comparative experiments, such as a study on ¹⁸F-FDG PET/CT lymphoma images[34], comparing seven semi-automatic segmentation methods and two fully automatic segmentation methods. The semi-automatic segmentation scheme based on SAC Bayesian method performed best in the Dice similarity coefficient within the observer and Dice similarity coefficient between observers, demonstrating strong stability and repeatability. Compared with manual segmentation, the method is more efficient and can also obtain similar lesion characteristics, which has the potential to replace manual segmentation and build high-quality large data sets, thereby promoting training and verification operations around deep learning models and improving model performance. When verifying the reliability of the segmentation method, the results of testing with large-scale data sets are more convincing. According to this understanding, researchers have advanced a larger-scale method evaluation process. In another study[33], 11 automatic PET segmentation methods were evaluated in 1223 lesions (median tumor and background contrast of 4.0). Two application scenarios were included for the segmentation effect of lesions with a median volume of 1.8 cm³, single lesion segmentation and overall disease burden quantification of lymphoma patients. The three-dimensional convolutional neural network DeepMedic performs the best in various evaluation indicators, and its consistency with the clustering algorithm and iterative threshold method in lesion segmentation is close to the physician's level. Driven by the actual demand for clinical applications, researchers began to focus on the impact of different segmentation methods on prognostic evaluation. Driessen et al[35] based on ¹⁸F-FDG PET/CT imaging data from 105 HL patients (35 newly diagnosed patients and 70 relapsed/refractory patients), compared the effects of the six segmentation methods in the accuracy, completeness, manual adjustment requirements and prognostic value of tumor metabolic volume, metabolic intensity, and disseminated radiomics characteristics of the six segmentation methods. The data from the study show that there is a high correlation between tumor metabolic volume, metabolic intensity, and most dissemination characteristics obtained by different methods and that the prediction effects are close. Although the SUV4.0 fixed threshold segmentation method may ignore small lesions with low FDG affinity, it requires minimal manual intervention, which is of great significance for future research and clinical applications.

This section compares the performance differences of different segmentation methods. The current model evaluation methods mainly rely on comparison with expert marking results. There is no unified evaluation standard. In addition, when performing performance comparisons between different methods, it is often based on different databases, and it is impossible to make direct comparisons quickly. In the future, it is recommended to establish a standardized evaluation system and public test data sets, and create more comprehensive performance evaluation indicators.

Identification and processing of factors affecting automatic segmentation

In normal biological distribution, ¹⁸F-FDG has a higher metabolic rate in the brain, heart and other parts or the rapid excretion of radiotracers in the kidneys, bladder and other parts will show significant physiological activity. However, in daily imaging diagnosis, this normal distribution pattern can significantly affect the performance of intensity-based PET image lesion detection and segmentation methods[36,37]. Therefore, some studies used CT information in ¹⁸F-FDG PET/CT imaging and combined with automatic detection technology to distinguish lymphoma from other highly metabolic sites.

The identification of normal FDG excretion and physiological absorption areas in the human body helps distinguish lesions from normal tissues. In recent years, deep learning methods have shown significant advantages in medical imaging analysis, especially in the recognition of complex physiological absorption patterns. Weisman et al[38] proposed a DeepMedic model based on 3D architecture. This model can implicitly learn the characteristics of high-uptake normal tissue areas during training, and its average detection true positive rate reaches 85%. Due to higher heterogeneity in physiological uptake in areas below the diaphragm (such as the bladder, kidney, and ureter) the model performed better below the diaphragm than above it. At the same time, in a study by Bi et al[36], a method based on multi-scale superpixel encoding, class-driven feature selection and classification model was used to classify the ¹⁸F-FDG PET/CT images of 40 lymphoma patients. The results showed that the classification accuracy of this method was increased by about 5.87% on average compared with the traditional method. In addition, the study also used the receiver operating characteristic (ROC) curve to evaluate the classification effect of FDG excretion and physiological absorption areas at different superpixel scales, verifying the effectiveness and advantages of the proposed method. In addition, the texture features in ¹⁸F-FDG PET/CT imaging can also be used to distinguish lymphoma from other high-uptake lesions. In one study, Lartizien et al[39] applied support vector machines combined with a feature selection strategy based on six filtering methods to classify 156 lymphoma regions and 32 regions of interest with suspected but non-lymphomas. The multimodal imaging-assisted diagnostic system combining PET and CT characteristics achieved optimal performance in distinguishing cancerous from benign lesions, with an optimal area under the curve (AUC) of 0.911, indicating that multimodal imaging significantly improved diagnostic performance.

Inflammatory diseases and infectious processes are a major confounding factor affecting lymph node segmentation. They can cause increased uptake of FDG, which is thus confused with malignant lesions. Image acquisition parameters and reconstruction methods are also key conditions that affect the performance of automatic segmentation. Various PET scanner models, reconstruction algorithms (such as ordered subset expectation maximization, maximum likelihood expectation maximization) and post-reconstruction filter parameters will significantly affect image features, which in turn will have an impact on segmentation accuracy[40,41]. The European Nuclear Medicine Association Research Ltd. (EARL) guidelines provide standardized protocols for PET/CT imaging, which are key elements in ensuring consistent segmentation performance in multicenter studies[42]. In the segmentation of thoracic and abdominal lymph node lesions, respiratory movement artifacts pose a major challenge. Respiration may trigger a blur effect when patients undergo PET acquisition, causing the apparent size and shape of lymph nodes to change, especially lymph nodes near the diaphragm[43]. In addition to the traditional gating reconstruction technology, the motion-corrected image reconstruction (MCIR) algorithm based on respiratory motion correction has become a new research focus. Meng et al[44] showed that the MCIR algorithm integrates activity-attenuation matching and non-rigid registration into the image reconstruction process, which can significantly enhance the quantitative accuracy level of PET images and reduce motion artifacts.

Using methods such as CT information, deep learning technology and multimodal imaging analysis, significant progress has been made in distinguishing lymphoma lesions from normal physiological high uptake areas in ¹⁸F-FDG PET/CT imaging. However, the current automatic detection and segmentation methods still face many challenges: The increased FDG uptake caused by inflammatory diseases and infection processes may be confused with malignant lesions; various imaging parameters, reconstruction algorithms and post-processing methods will significantly affect the image characteristics and segmentation accuracy; artifacts caused by respiratory movements are particularly obvious in the chest and abdomen area and will change the apparent morphology of the lesions. Future research needs to focus on developing more robust differential diagnostic models, integrating standardized imaging protocols (such as EARL guidelines), improving motion correction technology. In addition, building a comprehensive analysis framework that can dynamically identify and distinguish high uptake patterns caused by different physiological, pathological and technical factors to further enhance the performance of automatic lymphoma detection and segmentation in terms of accuracy and reliability.

APPLICATION OF DEEP LEARNING IN DIAGNOSIS OF LYMPH NODE METASTASIS IN MULTIPLE TYPES OF CANCER

When exploring the progress and prognosis of cancer, lymph node metastasis is an important element of disease progression. It not only reflects the spreading ability of cancer cells, but also affects the survival rate and the selection of treatment plans. PET/CT, as a key imaging diagnosis tool, uses deep learning and AI technology to greatly improve the accuracy and efficiency of diagnosis (Table 1)[6,45-49].

Table 1 Summary of deep learning-based studies for lymph node disease diagnosis using 2-deoxy-2-fluorodeoxyglucose positron emission tomography/computed tomography imaging.

Ref.	No. of patients	Disease	Technical methods	Performance metrics
Chen et al[45] 2019	59	Lymph node metastasis of head and neck cancer	Combining a hybrid model of MaO-radiomics and 3D-CNN	Accuracy = 0.88; Macro-average = 0.89; mean-OVA-AUC1 = 0.95; multi-class AUC2 = 0.95
Yang et al[6] 2023	165	Distinguish between lymph node metastases in swollen lymph nodes in the neck and lymphoma involvement	DL-CNN, DL-SVM and a combined model	DL-CNN: Best model: ResNet50; AUC=0.845; Accuracy = 78.13%, DL-SVM: Best model: ResNet50; AUC = 0.901; accuracy = 86.96%, Combination model: AUC = 0.948; accuracy = 84.00%; sensitivity = 100.00%; specificity = 75.00%
Zhang et al[46] 2023	689	Lymph node metastasis of ESCC	AI-CAD	Doctors vs AI-CAD: (1) Accuracy: 0.712→0.833; and (2) Specificity: 0.697→0.891. Diagnostic results in 12.1% of patients were corrected with AI-CAD assistance
Qiao et al[47] 2022	228	Lymph node metastases of NSCLC	Radiomics nomogram based on 18F-FDG PET/CT	Training set: AUC = 0.884; test set: AUC = 0.881
Trägårdh et al[48] 2022	660	Local recurrence of prostate cancer, lymph node metastasis and bone metastasis	UNet3D CNN model	Doctor sensitivity vs UNet3D CNN model sensitivity: (1) Local recurrence: 78% vs 79%; (2) Lymph node metastasis: 78% vs 79%; and (3) Bone metastasis: 59% vs 62%
Trägårdh et al[49] 2022	221	Pelvic lymph node metastasis of prostate cancer	CNN model	Doctor average sensitivity vs CNN model sensitivity = 77% vs 82%

¹Mean- one-vs-all- area under the curve is defined as the average of the one-versus-all area under the curve, which can be used as a measure of how well the classifier separates each class from all the other classes.

²Multi-class area under the curve is an extended definition of two-class area under the curve that averages pairwise comparisons to evaluate multi-class classification problems.

Mean-OVA-AUC: Mean one-vs-all area under the curve; ESCC: Esophageal squamous cell carcinoma; NSCLC: Non-small cell lung cancer; CNN: Convolutional neural network; AI-CAD: Artificial intelligence-based computer-aided diagnosis system; ¹⁸F-FDG PET/CT: 2-deoxy-2-fluorodeoxyglucose positron emission tomography/computed tomography.

Although lymph node metastasis patterns may differ in different cancer types, their importance in cancer research is clear. It can be seen that a recent study[45] proposed a fusion method based on evidence inference, integrating multi-objective radiomics characteristics with the prediction results of 3D-CNN to improve the accuracy of the detection and prediction of lymph node metastasis in head and neck cancer. Whether it is simply using CT images, PET images, or fusion of two imaging modalities, the designed 3D-CNN combined with multi-objective radiomics model shows better performance than XmasNet and traditional radiomics methods. After using the evidence reasoning strategy, the hybrid model that fuses multi-objective radiomics and 3D-CNN has obvious advantages over a single model. When distinguishing between nodes in various lymph node states, this model shows excellent diagnostic performance. The AUC of normal and suspicious nodes is 0.96, the AUC of normal and involved nodes is 0.97, and the AUC of suspicious and involved nodes is 0.91. Further analysis revealed that by controlling the 3D-CNN weight between 0.6 and 0.8, the model performance can reach the best level. When the weight value exceeds this range, it will cause the model performance to decline.

Accurate identification of lymph node metastasis at the same anatomical location but originates from different diseases is one of the important challenges facing the current medical field. Yang et al[6] developed a PET/CT imaging-assisted diagnosis system based on deep learning, which is used to distinguish malignant tumor lymph node metastasis and lymphoma lesions, and has achieved good results in the diagnosis of cervical lymph node swelling. The researchers used two technical routes to carry out related research. One of the ways is to construct a deep convolutional neural network model according to the ResNet50 architecture, which successfully distinguishes the category of cervical lymphoid tissue lesions. For the test set in PET/CT mode, the AUC reached 0.845. Another solution is to build a deep learning support vector machine model based on ResNet50. The diagnostic effect of this model has improved significantly. The AUC has reached 0.901 after detection. Deep learning and manual extracted radiomics features are combined into the support vector machine model. The model has achieved further optimization and the AUC has been improved to 0.948.

In the diagnosis and treatment process of ESCC, it is of great significance to accurately diagnose whether lymph nodes have metastasis. Recent research[46] has developed an AI-based computer-aided diagnosis system (AI-CAD), which uses PET/CT images to predict lymph node metastasis status in ESCC patients. In a multicenter study of 689 patients, the prediction accuracy of the AI-CAD system reached 74.4% in the external verification cohort, and it reached a good match with the diagnosis of clinical experts (kappa = 0.674 and 0.587, P < 0.001). More importantly, with the assistance of AI-CAD, the accuracy of clinical experts in diagnosing lymph node metastasis increased from 71.2% to 83.3%, and the specificity increased from the original 69.7% to 89.1%. These findings confirm that the AI-CAD system serves as an effective aid to enhance the accuracy of preoperative lymph node metastasis diagnosis in patients with ESCC.

In order to improve the accuracy of predicting lymph node metastasis in occult state in patients with non-small cell lung cancer (NSCLC), Qiao et al[47] conducted a radiomics study based on PET/CT. In this study, 228 surgically confirmed NSCLC patients were included, and these patients were divided into training sets (159 cases) and verification sets (69 cases). The research team adopted a standardized radiomics analysis process, including image segmentation using ITKsnap3.8.0, feature extraction using AK 3.2.0, and then feature screening with Python 3.7.0. Based on the logistic regression algorithm, the researchers developed a prediction model that combines 6 key radiomics features and tumor position elements (center or peripheral). The model demonstrates excellent predictive performance in the training set and the validation set, with the AUC of 0.884 (95%CI: 0.826-0.941) and 0.881 (95%CI: 0.8031-0.959), respectively. The clinical practical value of this model is further verified with the help of clinical decision curve analysis.

Trägårdh et al[48] is committed to developing and validating an automated system based on AI technology to identify and quantify primary prostate tumors, lymph node metastasis, and bone metastasis in ¹⁸F-PSMA1007 PET/CT imaging scans. The study included imaging data from 660 patients for analysis. The AI system achieved an average sensitivity of 79% when identifying primary tumors or recurrent lesions of the prostate, and the sensitivity to lymph node metastasis lesions was also 79%, while the sensitivity in the detection of bone metastasis lesions was 62%. In contrast, nuclear medicine specialists have sensitivity to these three types of lesions: 78%, 78%, and 59%, respectively. In addition, Trägårdh et al[49] developed an AI algorithm based on 211 cases of ¹⁸FPSMA PET/CT imaging data to identify pelvic lymph node metastasis in high-risk prostate cancer patients. The detection rate of this algorithm reached 82% and exceeded the average level of 77% for clinicians, and the mean of false positives per subject was 1.8. The conclusions show that the AI system has the same diagnostic capabilities as professionals while maintaining an acceptable false positive rate.

STUDY ON DIFFERENTIAL DIAGNOSIS OF LYMPH NODE LESIONS ASSISTED BY DEEP LEARNING

Lymphoma is a type of hematologic malignant lesions with many subtypes. It can generally be divided into HL and the so-called NHL[50]. The clinical manifestations of NHL are closely linked to the subtype. Inert lymphoma develops in a slow progressive manner. For example, follicular lymphoma (FL) accounts for 20%-25% of all NHLs[51], and the onset is occult and often manifests as painless systemic lymph node enlargement. About 30%-40% of patients do not show symptoms in the early stage of the disease, and some may experience hematocrit due to bone marrow involvement[51]. About 2%-3% of cases may turn into aggressive lymphoma[52], and invasive subtypes are acutely accompanied by systemic symptoms or abnormal organ function. For example, DLBCL accounts for about 30%-40% of all NHLs and is the most common subtype[53], with rapid growth in clinical manifestations, B symptoms (fever, night sweats, weight loss), and organ-specific symptoms (such as gastrointestinal bleeding, central nervous system involvement)[54,55], and about 40% of patients will develop drug resistance or recurrence in first-line treatment[54]. The typical clinical manifestations of HL are mainly painless lymph node enlargement, which is more common in lymph nodes in the neck, supraclavicular or mediastinum. Patients often have fever and night sweats, and are accompanied by systemic symptoms of weight loss of > 10% within 6 months. In the late stage, the spleen, liver, bone marrow and other parts may be involved, showing corresponding symptoms (such as spleen enlargement and hematocrit)[56]. The metastasis pattern of solid tumor lymph nodes is different in essence from lymph node infiltration in hematologic diseases. The latter is mostly shown to be progressive destruction of lymph node structure rather than simply forming metastatic lesions.

In today's medical diagnosis and treatment, the diagnosis of most hematopoietic and lymphoid tissue malignant tumors mainly relies on pathological examinations, and surgical procedures are generally used to obtain lesion tissue. This diagnostic method has certain limitations, just like tissue sample collection may not be sufficient, or the subjective judgment of the pathologist during the diagnosis process may affect the results, and the complex classification and diagnostic criteria of lymphoma itself are also prone to misdiagnosis. Technical problems associated with molecular gene detection may cause deviations[57]. It is more worth noting that pathological examinations are invasive operations, perhaps because they cannot fully reflect the overall situation of recurrent or refractory cases, thus having a potential impact on the choice of subsequent treatment plans. A large number of research today focus on exploring the use of AI models in ¹⁸F-FDG PET/CT image analysis, especially for distinguishing lymphoma from other malignant tumors and lymphomas of different subtypes.

In identifying different types of lymphomas, researchers conducted a series of exploratory studies based on AI, and achieved significant results. In the retrospective study performed by de Jesus et al[58], 120 patients were included, of which 44 were diagnosed with FL, and the remaining 76 were diagnosed with DLBCL. The study used an integrated classifier based on logistic regression and tree models (including AdaBoosting, Gradient Boosting, and XG Boosting) to analyze selected radiomics features and simultaneously compare their performance with maximum standardized uptake value (SUVmax)-based logistic regression models. The radiomics-based Gradient Boosting model performs better than the traditional SUVmax model in distinguishing two lymphoma types. The AUC of the Gradient Boosting model reaches 0.86, while the AUC of the SUVmax model is 0.79, and its accuracy is 80% and 70% in turn.

In addition to the differentiation of lymphoma subtypes, differential diagnosis of lymphoma and other diseases is also of great clinical significance. Researchers have also made in-depth explorations in this field. Lovinfosse et al[59] assessed the diagnostic consistency of lymphoma and sarcoidosis and HL and DLBCL. The consistency of the observer when distinguishing lymphoma from sarcoidosis was moderate, with a Fleiss k value of 0.66. Using the best machine learning model, the radiomics characteristics and age factors of texture analysis (TLR) were combined, and the AUC values obtained from the differential diagnosis of lymphoma and sarcoidosis were 0.94 and 0.85. When distinguishing HL from DLBCL, the AUC of the TLR radiomics model based on the lesion was 0.95 and the AUC of the patient-based model (combining the original radiomics characteristics and age) was 0.86. Machine learning-based imaging numerology technology can effectively analyze the lesion characteristics of sarcoidosis and lymphoma and has excellent diagnostic efficacy, even exceeding the doctor's level in some cases. In the diagnosis process of doctors, there are still obvious differences between observers, which further highlights the stability and potential value of machine learning methods.

In the specific disease category of primary gastric lymphoma (PGL), researchers have also worked hard to explore the application value of ¹⁸F-FDG PET/CT imaging characteristics in differential diagnosis. DLBCL and mucosa-associated lymphoid tissue (MALT) lymphoma are the two most prevalent subtypes of PGL, but the two show significant differences in clinical characteristics. Albano et al[60] discussed the value of ¹⁸F-FDG PET/CT and its radiomics characteristics in predicting the final diagnosis of PGL patients. A total of 91 newly diagnosed patients with PGL were included in this study, and all patients underwent ¹⁸F-FDG PET/CT. Study data showed that 83 (90%) patients had positive FDG uptake, of which 54 were DLBCL. PET/CT metabolic characteristics of DLBCL, such as disease stage and tumor size, are significantly higher than those of MALT, and Helicobacter pylori infection is more common in patients with MALT. This study shows that the ¹⁸F-FDG PET/CT parameters can distinguish DLBCL and MALT with high accuracy, providing an important reference for clinical diagnosis.

Research on the application of AI in ¹⁸F-FDG PET/CT image analysis has shown significant progress. Some studies have found that the accuracy of AI model diagnosis even exceeds the diagnosis level of clinicians. At this stage, there are still problems such as insufficient sample size, lack of large-scale prospective verification processes, and data standardization to be dealt with. With the continuous development of technology and the continuous advancement of clinical practice, relying on larger-scale multicenter verification research, improving the accuracy and interpretable effects of algorithms, and promoting the standardized use of diagnostic systems, AI will probably become an important auxiliary device in the field of lymphoma diagnosis, delivering more accurate and efficient diagnostic support for clinical practice.

APPLICATION OF DEEP LEARNING IN THE EVALUATION OF LYMPHOMA EFFICACY

The rate of complete response to diffuse and aggressive NHL in adults is higher, usually between 60 and 80%, however, one of the major challenges in treating the disease is that about 20 to 40% of patients may experience relapse. HL is more sensitive to chemotherapy. Generally, most patients are curable, but 5% to 10% of patients still have initial refractory disease, and 10% to 30% of patients may relapse later[61,62]. The Deauville five-point system proposed in 2009 is a visual evaluation method based on the relationship between FDG uptake value and mediastinal blood pool and liver. It is highly reproducible and aims to be used to interpret PET/CT images of lymphoma. This scoring system has been widely used in early treatment response evaluation of lymphoma and has shown good reliability in interobserver consistency[63-65]. In addition, other prognostic tools for risk stratifying the management of lymphoma patients include the international prognostic index (IPI) and its optimized versions (such as R-IPI and NCCN-IPI), TMTV, and tumor maximum diameter (Dmax), etc.[21,66]. Because NHL contains multiple subtypes and the biological characteristics and FDG uptake levels of different subtypes vary greatly, and the current treatment strategies for each subtype are also different, even if prognostic indicators such as IPI are used, there are still major challenges in evaluating prognostic effects.

DLBCL, as the most common type of NHL in the world, most studies are currently focusing on DLBCL treatment. To improve the early identification of the risk of recurrence after first-line treatment of DLBCL, in one study[67] researchers included 271 DLBCL patients and constructed a comprehensive predictive model based on ¹⁸F-FDG PET/CT imaging and combined with clinical features to identify patients with a higher probability of progression or recurrence after first-line treatment. The model filtered the best features and classifiers through the cross-combination method, achieving predictive performance better than IPI and Deauville scores in the test set (AUC reaches 0.898), while also showing better performance in predicting PFS (C-index 0.853). At the same time, in exploring the prognosis prediction of DLBCL patients, Frood et al[68] collected 229 DLBCL patients who met the enrollment criteria and 62 of them (27%) had disease progression within two years. Comprehensive model combination of ¹⁸F-FDG PET/CT imaging characteristics and clinical data predicted that the efficacy of PFS in patients with DLBCL two years after R-CHOP chemotherapy was better than that of MTV-only model (AUC = 0.67), and showed good predictive effects in independent test sets. A recent study[69] is based on the imaging data of mid-term ¹⁸F-FDG PET/CT, and is determined to build a multimodal deep learning model to estimate the possible treatment failure of patients with low-risk DLBCL. This study included 205 DLBCL patients undergoing ¹⁸F-FDG PET/CT scans and first-line standard treatment in the primary dataset for model development, and 44 additional patients were included in the external dataset for model extension evaluation. The final model optimized by comparison targets, named the contrast blended learning model, performed best in the main dataset with an accuracy of 91.22%, while the AUC reached 0.926. In the external dataset, the accuracy and AUC stayed at 88.64% and 0.925 respectively, confirming that the model has good generalization ability. On the occasion of the study of the special subtype of primary gastrointestinal DLBCL, Jiang et al[70] constructed a comprehensive model based on radiomics characteristics, metabolic parameters and clinical risk factors to predict PFS and OS after treatment with R-CHOP regimen in primary gastrointestinal DLBCL patients. The model showed good prediction results in the training set (PFS: 0.825, OS: 0.834) and the verification set (PFS: 0.831, OS: 0.877). The decision curve analysis further affirmed the clinical application value of this model when predicting PFS and OS.

In response to the prognostic evaluation of a novel therapeutic approach to CAR-T cell therapy, Zhou et al[71] evaluated the effect of CAR-T cell treatment on PFS and OS in patients with relapsed or refractory DLBCL. Sixty-one DLBCL patients who underwent ¹⁸F-FDG PET/CT examination before CAR-T cell infusion were included and randomly divided into training sets (n = 42) and validation sets (n = 19). The prediction effect of radiomics combined clinical model is better than that of clinical models alone, with the AUCs of PFS in the training set being 0.776 and 0.712, respectively, while the AUCs of OS are 0.828 and 0.728, respectively. In the verification set, the AUCs of PFS are 0.886 and 0.635, respectively, while the AUCs of OS are 0.778 and 0.705. In another study, Ligero et al[72] compared the differences in PET-based radiomics characteristics with traditional PET biomarkers in predicting the therapeutic efficacy of CAR-T cell in patients with relapsed or refractory large B-cell lymphoma. The study included 93 large B-cell lymphoma patients who received CAR-T cell infusion between 2018 and 2021 and divided into training sets (73 cases) and test sets (20 cases). The predicted efficacy of radiomics characteristics (AUC of 0.73) was significantly better than traditional biomarkers such as TMTV (AUC of 0.66) and SUVmax (AUC of 0.59). Further analysis shows that higher radiomics scores are closely related to longer PFS and OS.

In the prognosis of classical HL, Driessen et al[73] studied how to predict PFS in patients with relapsed or primary refractory classical HL after remedial chemotherapy and autologous stem cell transplantation. The study combined ¹⁸F-FDG PET/CT radiomics characteristics and clinical data from 113 patients to develop a machine learning-based predictive model for screening high-risk patients with low PFS within 3 years. The AUC of this model in the training data is 0.810 and the AUC in the verification data reaches 0.750.

This subsection shows the latest advancement of the prognostic evaluation system for invasive lymphoma (NHL/HL). As for the high-proliferation subtype of DLBCL, a large number of studies have integrated radiomics characteristics and clinical parameters to develop new prediction models that break through the limitations of traditional risk stratification. In special subtypes such as gastrointestinal primary and novel treatments, cross-modal prediction models also show excellent clinical value. However, the main problems are that the high heterogeneity of NHL leads to the limitation of model generalization ability, most of the existing studies are single-center retrospective analysis, and traditional metabolic parameters are not sensitive to microresidues. In the future, we will use multicenter prospective cohorts to verify the universal applicability of the optimization model, develop enhanced algorithms that combine molecular characteristics of tumor microenvironment, and use AI technology to dynamically monitor the treatment response, and finally establish a full-course management model based on precise typing.

CONCLUSION

This paper reviews the research progress of deep learning technology in lymph node detection, segmentation and clinical application. In the automatic lymphoma segmentation, the deep learning model reflects better efficiency and consistency than the traditional manual segmentation method. It can efficiently quantify MTV, TMTV, TLG, SUVmean and other indicators to assist clinical decision-making[18-25]. In radiomics, lymph node metastasis is often distributed in a wide range and is easily missed. A significant advantage of fully automatic AI models is that they can effectively improve workflow efficiency and avoid the complex task of nuclear medicine doctors identifying lesions one by one[28,30,32]. The semi-automatic segmentation also showed good results in handling false positive results and complex cases[26,27]. Deep learning technology can also effectively identify FDG excretion and physiological absorption areas, thereby reducing the occurrence of false positives and improving the accuracy of lymph node detection[6,36-38].

In the treatment effect evaluation stage, deep learning models can be used to predict patients' treatment response and prognostic outcomes, predict PFS and OS in patients with DLBCL, and assist in the implementation of risk stratification[70,71]. Some studies also show that deep learning models are superior to traditional IPI and Deauville scoring systems in terms of predictive capabilities[67], and have achieved significant results in predicting the risk of treatment failure[69] and evaluating the therapeutic effect of CAR-T cells[71,72].

Deep learning also shows great potential in the detection and classification of lymph node metastasis in various types of cancer. For example, the PET/CT imaging-assisted diagnostic system based on deep learning can effectively distinguish malignant tumor metastasis and lymphoma-related lesions, while assisting in the accurate diagnosis of lymph node metastasis[6]. The combination of multi-objective radiomics and 3D-CNN and other methods such as evidence reasoning further improve the prediction accuracy[45].

Even though deep learning technology shows great potential in the application of ¹⁸F-FDG PET/CT for the diagnosis of lymph node disease, it still faces many challenges before achieving full clinical application. Model generalization ability is the primary challenge. The poor performance of fully automatic models in dealing with rare cases may be related to the limited training data. The high heterogeneity of NHL will further hinder the generalization ability of the model, and there are obvious differences in the biological characteristics and FDG uptake of different subtypes. To solve this problem, future research can focus on setting up larger and more diverse data sets to cover various rare subtypes, and at the same time, building models with stronger generalization capabilities based on the data sets. The standardization of data is also of great significance. At this stage, the research mostly uses single-center retrospective data, and there is no unified data collection and preprocessing standards. The differences in equipment, changes in imaging protocols and uneven image quality in different centers seriously affect the repeatability and reliability of the model. This requires the establishment of unified data acquisition standards and quality evaluation tools in the future, and a multicenter collaboration network can be established to strengthen the connection between each center. The integration of AI and existing clinical decision support systems still faces obstacles to technology and practice. Clinicians still need to improve their acceptance of AI diagnostic results. In the future, it can optimize human-computer interaction interfaces and promote the development of interpretable AI. In the application of AI medical applications, ethical and regulatory considerations are becoming increasingly critical, involving patient informed consent, data privacy protection, and the definition of responsibility. A complete legal framework and technical specifications must be formulated. Traditional metabolic parameters still have limitations in monitoring micro-residual lesions, and more sensitive detection algorithms are needed in the future. By adopting a systematic approach to addressing these challenges, the AI-powered ¹⁸F-FDG PET/CT lymph node disease diagnosis technology will serve clinical practice more effectively and ultimately benefit more patients.