Application of artificial intelligence and radiomics in the prediction of lymph node metastasis and tumour grading of oral cancer - a systematic review and meta analysis.

Introduction
Head and neck cancers (HNC) rank as the seventh most prevalent cancer across the globe, predominantly including oropharyngeal squamous cell carcinoma (OPSCC) [1]. Assessing the lymph node (LN) status determines the staging, management, and ultimately the survival outcomes of patients with OSCC. Cervical lymph node metastasis (LNM) is associated with poor prognosis and is one of the most important independent prognostic factors in OSCC [2]. The dissection of metastatic LNs at the time of resection of the primary tumour can significantly reduce the regional recurrence rate and enhance the survival of patients with OSCC [3]. Computed tomography (CT) and magnetic resonance imaging (MRI) are widely used to evaluate the status of cervical LNs in patients with OSCC [4, 5]. Besides the intrinsic challenges of identifying LNM, such as tissue properties and technical limitations, the paramount factor in precise diagnosis is human error influenced by professional expertise and the demanding workflow of radiologists [6, 7]. Computerized assistance diagnostic techniques that amalgamate qualitative and quantitative imaging characteristics to identify lymph node metastasis [4]. The radiological feature-based AI plays an essential role in tumour diagnosis, staging, and predicting the treatment response and prognosis, demonstrating its potential as a non-invasive auxiliary tool for personalized medicine [8].
Artificial intelligence (AI) has recently been applied to the evaluation of radiology images, as it excels at automatically recognizing complex patterns in imaging data and providing quantitative, rather than qualitative, assessments of radiographic characteristics [9]. ML involves the application of algorithms to automate decision-making processes using models that have not been manually programmed but have been trained on data. Artificial neural networks (ANN), a part of ML, aim to simulate the structure and function of the human brain. On the other hand, DL uses multiple layers of interconnected neurons of a convolutional neural network (CNN) capable of extracting depth features, thus enabling the processing and analysis of large and complex databases [10, 11].
Unlike deep learning, which employs intricate networks to extract and analyze features autonomously, in machine learning, the handcrafted features are derived from formulas centred on intensity histograms, shape characteristics, and texture matrices, which can identify phenotypical characteristics of radiological or histological images [12]. Delta radiomics examines temporal features and alterations to forecast a patient’s treatment responses [13]. Radiomics consists of extracting hundreds of quantitative features through automated or semi-automated software [14]. The traditional “radiomic workflow” comprises a sequence of procedures for reliable and uniform image data extraction. The steps encompass picture acquisition, the extraction of characteristics, and selection of features, which may be achievable using deep-learning (DL) radiomics, handmade radiomics, and delta radiomics [15].
The first order characteristic features include shape features, representing the shape and geometry of the region of interest (ROI), tumour volume, length axis ratio, surface area/volume ratio, etc [5]. These are ideally standardized for image characteristics in resolution, reconstruction and acquisition parameters, and clinical characteristics such as tumour stage, tumour classification or prognosis [16]. These features analyze the distribution pattern of voxel frequencies independently of spatial relationships, including metrics such as mean, standard deviation, and maximum voxel strength [17]. Second-order characteristics, or texture attributes, are employed to examine the spatial distribution of voxel intensity among voxels and can quantify heterogeneity within tumours. Examples include the co-occurrence matrix, which assesses continuous voxels with identical magnitude in a fixed direction. The neighbourhood grey-level different matrix evaluates the variance between the quantized voxel intensity and the mean intensity of neighbouring voxels within a defined distance [18].
Radiomics and artificial intelligence (AI) have emerged as prominent concepts in oncology studies, based on the potential of ML techniques for analyzing image attributes and extracting imperceptible data not visually discernible [19]. ML and DL systems are utilized to identify patterns in images, identify biomarkers, and integrate non-imaging elements that exceed human cognitive capacity [20, 21].
The present study aims to systematically review the published data on the performance of CT, PET-CT, MRI, and histopathology-based AI algorithms for predicting LN metastases and histopathology grading of OSCC. To elaborate on AI’s prediction performance and increase the evidence reliability, meta-analysis is performed by grouping studies according to the type of AI algorithm, image modality and assessment of the target sites.

Materials and methods

Search of databases
The present meta-analysis database search, eligibility criteria, and reporting adhered to the PRISMA protocol (PROSPERO Registration Number - CRD420251063771) [22]. An extensive search was performed across PubMed, Scopus, and Wiley Online Library. Furthermore, Google Scholar and the citations of pertinent papers were examined to gather further relevant studies.

Keywords and timeframe
Relevant publications were obtained through the following combination of keywords: (“Artificial Intelligence” OR “Deep Learning” OR “Machine Learning”) AND (“Oral Squamous Cell Carcinoma” OR “Oral Cancer” OR “Lymph Node Metastasis” OR Tumour Staging”) AND (‘Radiomics’ OR “Neural Network” OR “Convolutional Neural Network” OR “Deep Learning-based Radiomics” OR “Machine Learning-based Radiomics”) from these databases from January 2000 to May 2025.

Criteria for inclusion and exclusion
All research that integrated radiomics with any subfield of AI was included for grading or LNM prediction purposes in OSCC/OPSCC. The retrieved studies were evaluated for inclusion according to the PICO criteria.
Population: Research investigating the utilization of ML or DL on radiomic characteristics derived from CT, PET, MRI, ultrasound and histopathological imaging for tumour grading or staging purposes.
Intervention: ML/DL-based radiomics framework for prediction of tumour grading and lymph node metastasis.
Comparison: Model performance depending on any of the commonly used ML/DL parameters.
Outcome: Examined the performance of AI models for predicting tumour grading or LNM, which aid in the diagnosis, prognosis, or both.
Studies were excluded if: (1) the available outcome data could not be retrieved; (2) Comments or expert opinions, editorials, systematic or narrative reviews, and non-English-language articles; (3) animal studies, case reports, abstracts, and conference proceedings; or (4) duplicate publications; (5) Research focus on cancer subsites apart from OSCC/OPSCC.

Analysis and screening of search results
All possibly relevant papers from the database searches were evaluated for selection criteria. After the elimination of duplicates and extraneous studies, two separate researchers conducted a screening of potentially pertinent papers. During the initial step, the titles and abstracts were examined. Then, a thorough full-text evaluation of the articles was conducted to ascertain their relevance to the study aim.

Documentation of qualifying studies
Two unbiased investigators deliberated on addressing potential differences in the selected papers. A third adjudicator or consensus session was convened to address conflicts and facilitate discussions as needed. The inter-observer reliability among these investigators was assessed employing the Kappa coefficient (k = 0.89). A data extraction template was used to encapsulate the findings of reviewed investigations, contingent upon the consensus of two independent investigators.

Results
The literature search found 664 eligible papers in the evaluated databases. After eliminating duplicates, 494 articles were assessed. Four hundred forty redundant studies were removed, and 54 were retrieved based on eligibility criteria. Due to four unretrieved studies, 50 full texts were reviewed. Finally, the present systematic review included 40 studies after removing 10 papers not centred on radiomic-based oral or oropharyngeal cancer tumour grading or staging, of which 33 were selected for the quantitative synthesis, as illustrated in Fig. 1. (Prisma flow chart).

Evaluation of the quality of the reviewed investigations
The risk of bias (RoB) for the built prediction framework in the qualifying studies was evaluated utilizing the QUADAS-2 evaluation tool [19]. QUADAS-2 evaluates the quality and relevance of those included studies, as depicted in Fig. 2. The RoB evaluation encompasses four domains (D1-D4) for Risk of Bias analysis and three domains (D5-D7) for applicability concerns. The characterization of these responses is as follows: a “low” indicates no or low RoB, a “high” suggests the potential for significant RoB, and “moderate” signifies ambiguity regarding application and RoB. The overall findings were low if all four areas were assessed as low. The overall RoB was deemed high if any single domain received a high score. The overall RoB was considered moderate if at least one domain region exhibited moderate risk, while the others were assessed as low risk.

Data extraction
The information extracted from each included study was the first author’s name, publication year, study location, article details including patient count, imaging modalities, radiomics methodology and results encompassing conclusions for nodal status documented in Table 1 and histopathology grading in Table 2.

Characteristics of pertinent research
All 40 articles considered in this review (Tables 1 and 2) have been published in English [23–62]. Thirteen studies in Japan [27, 28, 30, 33, 39, 40, 42, 43, 49, 55, 56, 58, 59], ten were undertaken in China [23–25, 29, 34–36, 41, 47, 60], five in the United States of America [37, 38, 44, 50, 51], three in Italy [26, 32, 46], one in Austria [31], three in India [52, 61, 62], one in Switzerland [45], one in Korea [54], one in Germany [57], and two was a multicenter investigation [48, 53]. All studies had low RoB, except two examined studies demonstrated an uncertain RoB (Table 2) [29, 48]. Commercial and freely available software utilized for feature extraction in radiomics comprises ITK-SNAP [46], LIFEX [39], MATLAB [28, 35], PyRadiomics [34, 41, 48, 53, 54], ADC maps [60], 3D slicer [46] and Python [33, 36, 39]. Likewise, numerous ML or DL frameworks were suggested for feature selection including Decision Tree (DT) [51], Logistic Regression (LR) [32, 34], Least Absolute Shrinkage and Selection Operator (LASSO) [28, 60], Many-objective (MaO) radiomics model [51], K-Nearest Neighbours (KNN) [24, 46], Naïve Bayes [41, 46, 59], Support Vector Machine (SVM) [27, 33, 36, 37, 39, 41, 47, 51], Principal Component Analysis [54], Random Forest (RF) model [41, 45, 47, 59], Boot strap Forest [27], Linear Discriminate Analysis (LDA) [31, 61], Neural tanh boost (NTB) [27], Mask R-CNN [29, 36], Neural Net work (NN U net) [30], Artificial Neural Network (ANN) [35], and CNN [25, 37, 38, 40, 42–44, 49, 51, 56, 62].

Meta-analysis of diagnostic accuracy
After an extensive review and extracting relevant data, studies that satisfied the inclusion criteria were integrated into a bivariate random effects diagnostic test accuracy (DTA) meta-analysis. The pooled values of the meta-analysis of the included studies under different categories were listed in Table 3. Among the 29 studies for the meta-analysis, the pooled sensitivity, specificity, Diagnostic Odds Ratio (DOR), positive likelihood ratio and negative likelihood ratio of AI models for the prediction of LN metastases were 0.86 (95% CI 0.80–0.90), 0.91 (95% CI 0.87–0.93), 56.58 (95% CI 21.68–91.48), 9.13 (95% CI 5.80–12.45.80.45), and 0.16 (95% CI 0.11–0.22) respectively, which is displayed in the forest plot (Fig. 3). The forest plots of sensitivity and specificity for AI models utilized models such as ML and DL algorithms, imaging methodologies such as CT, PET-CT, MRI and HP and the assessment sites involving tumour site and lymph node for the prediction of LNM were represented in Figs. 4, 5, 6, 7, 8, 9, 10 and 11, respectively. The forest plots of sensitivity and specificity for AI models for histopathological grade assessment of OSCC were illustrated in Fig. 12. The meta-analysis of AI models for the prediction of histopathological grading of OSCC, revealed the pooled sensitivity, specificity, DOR, positive likelihood ratio and negative likelihood ratio of 0.88 (95% CI 0.54–0.98), 0.82 (95% CI 0.76–0.87), 34.38 (95% CI 24.24–103), 8.71 (95% CI 5.0–22.41.0.41), and 0.25 (95% CI 0.14–0.37) respectively (Fig. 12). The SROC curve with a 95% confidence region and prediction region of sensitivity and specificity for lymphnode status assessment and histopathological grade assessment are illustrated in Figs. 13 (a-b), respectively.

Heterogeneity
Substantial heterogeneity was observed in sensitivity and specificity analysis for lymph node stage assessment, revealed by Var(logit(sen)) and Var(logit(spec)) values of 0.932 and 0.898, respectively (Table 3). Substantial heterogeneity was observed in the sensitivity analysis of histopathological grading of OSCC, depicted by Var(logit(spec)) value of 2.389, while there was minimal heterogeneity in specificity analysis, represented by Var(logit(sen)) value of 0.139 (Table 3).

Publication bias
Generalized Egger’s regression test was used to assess publication bias of study effects in the diagnostic accuracy reported by the AI models of each included study. The probability value of (< 0.05) confirmed a significant publication bias. In addition, the calculated false positive rate, logit (sens) and SE(E(logitSe)) values were displayed for each model category in Table 3.

Discussion
Besides understanding the extent of oral cancers, their degree of lymph node metastasis (LNM) is crucial for adequate malignancy care. Radiomic analyses may predict cervical LNM of oral cancer using preoperative radiographic diagnostic modalities. Radiomics is intended to aid clinical decision-making by extracting quantitative information from medical imaging [63–65]. The radiomics methodology demonstrated superior diagnostic efficacy in distinguishing benign from metastatic cervical lymph nodes compared to qualitative analysis based on standard CT criteria [39]. Ariji et al. utilized a DL system with segmented CT images of lymph nodes, yielding diagnostic outcomes compared to radiologists [30]. Ren et al. collected imaging traits from MRI scans and proved that the MRI radiomics could differentiate stage III and stage IV from stage I and stage II HNC [34].
Certain investigators have suggested hybrid (ML&DL) prediction models to enhance predictive accuracy [66, 67]. Chen et al. suggested a hybrid predictive framework that leverages radiomics using deep and machine learning approaches with spatial context-sensitive data [25]. The hybrid approach attains superior accuracy to XmasNet and Radiomics techniques for predicting lymph node metastasis using PET and CT. The Mask Region-based CNN, which is centred on the Faster Region-based CNN, specifically generates high-quality masks for multiple instances of an object to predict metastasis of cervical lymph nodes [29].
Research indicates that the T-stage of tumours, viral status, and LN involvement significantly affect cancer prognosis [68–70]. Kann et al. reported that a DL-based CNN can be effectively trained to detect nodal metastases and extranodal extension (ENE) with higher accuracy [71]. Wang et al. documented the application of radiomics integrated with ML to develop a T-staging model for carcinoma of the larynx for predicting the T-stage of the tumour with the highest precision. Most included studies aimed to construct an AI model utilizing CT images, capable of identifying, locating, and differentiating lymph nodes with greater precision, eventually replacing the manual techniques [72]. Appropriate tumour staging evaluation using radiomics before therapy can assist in choosing a treatment and diminish recurrence and unfavourable reactions.
ML-based MRI prediction models can differentiate between aggressive and non-aggressive papillary thyroid carcinoma before surgical therapies, hence enabling the development of personalized treatment strategies [73]. A radiomics model was developed to forecast LNMs in early-stage cervical cancer utilizing T2-weighted and contrast-enhanced T1-weighted imaging, which effectively predicted N staging of the tumour [74]. These exploratory investigations indicate that the radiomics prediction framework may be a non-invasive diagnostic tool for OSCC before treatment.
In clinical practice, the histopathologic grading of cancer is generally conducted following assessing histopathological images of lesional tissue sections. Numerous investigations highlighted the likelihood of PET/CT/MRI-derived radiomic characteristics to create a DL-based radiomics framework for oral cancer grading [32, 59, 60]. Das et al. integrated a radiomic basis function with kernel principal component analysis and employed a random forest classifier to predict the grading of OSCC tumours in distinguishing well-differentiated tumours from moderately or poorly differentiated ones. The clinical and histopathological image combined CNN model showed superior efficacy with sensitivity and accuracy of 98% for tumour grading [62]. Romeo et al. study reported that apparent diffusion coefficient-based radiomics may serve as a valuable and potentially non-invasive technique to predict the histologic grade of cancer of the tongue and the floor of the mouth [46]. These investigations primarily concentrate on pre-treatment grading, differentiation of tumours from inflammation and necrosis, predicting tumour status post-treatment, and predicting survival and adverse effects following treatment [5].
The prospective use of CT texture analysis in assessing pathological lymph nodes has been documented in head and neck cancer, exhibiting perfect accuracy. Cell and nuclear shapes have long been a key focus in histomorphometry analysis of various cancer subtypes, including oral cancer [75]. More recently, these analytical approaches have been extended to assess additional cell types, such as tumour-infiltrating lymphocytes (TILs) [76]. In mathematics, the fractal dimension helps to describe the complexity of the shape of the cells and the nucleus [77]. Tissues with higher nuclear fractal scores exhibit increased morphological complexity, whereas those with lower scores show less structural irregularity [78]. Bilgin et al. pioneered the proposal of spatial graph algorithms to analyze the spatial arrangement of cells in pathology images quantitatively [79]. The features are constructed within a framework that examines local differences in nuclear morphology, including shape, size, and orientation [80]. Lu et al. extended their analysis to include variations in the orientation of individual nuclei. Additionally, morphological features, such as area ratio, distance ratio, perimeter ratio, nuclear boundary smoothness, and Fourier descriptors, have been employed to characterize shape variations across different cell types, including cancer nuclei and tumour-infiltrating lymphocytes (TILs) [81].
The result of the present meta-analysis demonstrates the promising accuracy of radiomics models in predicting LNM in oral and oropharyngeal cancers, with a pooled AUC of 89%, 93%, and 81% for CT, PET-CT and MRI-based models, respectively. In addition, the present study observed a pooled AUC of 85% and 90% for machine learning and deep learning models, respectively. The present study also found a pooled AUC of 86% for histopathological grading prediction of OSCC, significantly surpassing conventional diagnostic methods. CT radiomics is widely available and offers high structural detail, but is sensitive to acquisition variability. PET/CT adds metabolic information, improving sensitivity, but is limited by lower resolution and standardized uptake value (SUV) variability. MRI provides superior soft-tissue contrast and functional data, though sequence heterogeneity hampers standardization. Deep learning achieves high accuracy with large datasets but lacks interpretability, while classical machine learning is more transparent and suitable for smaller datasets but is limited by feature stability [82].

Limitations
Challenges include difficulty interpreting the generated feature representations, the need for substantial and accurately labelled datasets to train deep learning networks, and uncertainties regarding how well these approaches perform across different testing sites. Nonetheless, the discipline is advancing towards greater consistency using different modelling methodologies. Additionally, services such as GitHub have significantly enhanced the efficiency of software and script release, sharing, and validation. Ultimately, ontologies have undertaken a deliberate initiative to improve data annotation and informatics norms across various data sources, facilitating model development and data transfer. These processes facilitate the development of effective models for clinical use, thus enabling the customization of therapies to enhance therapeutic outcomes and quality of life.

Conclusion
The selection of a modelling technique is contingent upon the nature of the information and the research objectives. Radiomic CT characteristics of OSCC provide information regarding tumour heterogeneity and may possess the capability to forecast histopathologic attributes. These exploratory investigations suggest that the radiomics prediction model may serve as an additional non-invasive diagnostic tool for OSCC before treatment, enhancing the objectivity and accuracy of tumour diagnosis, predicting tumour metastasis and providing guidance for future therapies.