Review on applications of metastatic lymph node based radiomic assessment in nasopharyngeal carcinoma

INTRODUCTION

Nasopharyngeal carcinoma (NPC) is cancer arising from the epithelium of the nasopharynx, frequently originating at the Fossa of Rosenmüller. According to World Cancer Research Fund International, NPC is ranked the twenty-second most common cancer worldwide, with over 133,000 new cases in 2020^[1]. Although NPC is uncommon in western countries, it is characterized by distinct geographical prevalence and male predominance in the orient. More than 80% of the global cases were diagnosed in endemic areas such as Southern China and Southeast Asia. For instance, in Hong Kong, it was the seventh most common cancer in men and the seventeenth most common cancer in women in 2019, with a male-to-female ratio of 2.9 to 1^[2].

The staging of NPC adheres to the Tumor, Node, Metastasis (TNM) system of the Union for International Cancer Control (UICC) and the eighth edition of the American Joint Committee on Cancer (AJCC)^[3]. For accurate disease staging, magnetic resonance imaging (MRI) or positron emission tomography with computed tomography (PET-CT) images are essential. The overall survival rate in early-stage (stage I or II) NPC can be as high as 93%, while it drops to 63% in the advanced stage (stage III to IVb)^[4]. A meta-analysis showed that 85% of locally advanced NPC patients presented with nodal metastasis at diagnosis and bear a poor survival outcome^[5].

With the recent trend of investigating the applications of artificial intelligence (AI) and deep learning (DL), the use of radiomics has drawn more attention in the precision therapy of NPC. There are numerous comprehensive reviews on the applications of AI in NPC^[6,7]. These reviews focused mainly on the radiomic analysis of the gross tumor volume^[8], while others evaluated the radiomics in PET-CT of NPC lymph nodes and primary tumors. A recent study by Xia et al. proposed that lymph node assessment as opposed to that of primary tumors performed better in survival prediction^[9]. However, to the best of our knowledge, minimal effort has been placed to explore and summarize radiomics-related studies on lymph node metastasis in NPC^[9].

This review aims to summarize the current evidence for the radiomics use for NPC with nodal metastasis and explore its clinical potential in the prognostic prediction, treatment response, and overall survival of NPC patients with lymph node metastasis.

ARTIFICIAL INTELLIGENCE AND RADIOMICS

The term “radiomics” is regarded as novel among clinicians. It is often misunderstood as synonymous and interchangeable with artificial intelligence, machine learning, and deep learning. This section provides an overview and the interrelationship of these terminologies.

Artificial intelligence (AI) is a constantly evolving technology, and with recent efforts exploring the applications of AI and medical imaging in NPC from diagnosis, imaging segmentation for radiotherapy planning to treatment monitoring and prognostication.

Ng et al. reviewed the implementation of AI in NPC, summarizing four main types of applications, namely auto-contouring, diagnosis, prognosis, and treatment planning^[7]. AI, first termed by an American computer scientist John McCarthy in 1956, has been expanded to a wide spectrum of applications and abilities demonstrated by the machine^[10]. For example, knowledge representation, natural language processing, automatic reasoning, machine learning, intelligent robots, and automatic programming are under the category of AI^[11].

Machine learning is a branch of AI using algorithmic systems to perform tasks such as prediction and diagnosis of interest^[12]. Deep learning, also called deep neural network learning, employs a neural network to allow automatic learning capability mimicking a learning process in the human brain^[13].

Radiomics is a quantitative approach in which models are built based on high-dimensional spatial features in medical imaging that are not visually perceptible and interpretable by human beings^[14,15]. Radiomics is a high-frequency lexicon in oncological medicine. It requires of the acquisition, preprocessing, and extraction of a huge amount of data in medical images for quantitative analysis, providing additional information based on the differences in tumor characteristics such as shape and textures^[16]. In radiomics, a large number of features are generated to determine the non-linear relationship between them. With AI development, computer scientists adopt deep learning algorithms and machine learning methods for tumor segmentation, feature extraction, and selection, and finally formulating a model for analysis [Figure 1].

Figure 1. Relationship of artificial intelligence, machine learning, deep learning, radiomics.

METHODS

A literature search was conducted to retrieve all studies of NPC with nodal metastasis under radiomic analysis. Keywords were combined using Boolean logic to produce the resulting search phrase: (“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural networks, computer” OR “algorithms”) AND (“nasopharyngeal neoplasm” OR “Nasopharyngeal carcinoma” OR “NPC”) AND ("lymph nodes” OR “lymphatic metastasis”). Using this search phrase, studies were searched on PubMed, Ovid Medline, and Embase between Jan 2002 to June 2022. The studies were screened, with duplicates removed. Inclusion and exclusion criteria were defined to allow evaluation only on those applying radiomics in NPC with nodal metastasis [Table 1]. The authors looked into the methodology and image segmentation methods of individual studies. Studies with lymph node-based radiomic analysis as regions of interest (ROI) were included. Studies that are limited to primary tumors as ROI or those without detailed descriptions of lymph node analysis were excluded. Fourteen studies were selected to highlight the current state-of-the-art lymph node-based radiomics applications in NPC.

RADIOMIC STUDIES IN LYMPH NODE ANALYSIS AND RISK STRATIFICATION

There are studies exploring the full potential of radiomic analysis in both primary tumor and nodal metastasis in head and neck cancer, suggesting that the additional metastatic lymph node radiomic information could be a complementary prognostic factor in loco-regional control of the disease^[17,18]. Studies extracted in Table 2 investigated the radiomic features of lymph node metastasis in NPC patients to predict high-risk patients requiring treatment adjustment and modification.

Five studies from 2018 to 2022 adopted various machine learning algorithms for radiomic analysis with regions of interest in nodal metastasis. All these studies achieved AUCs between 0.725-0.984 in training and testing sets. Of these, four studies employed MRI images for analysis^[19-22]. Li et al. identified the NPC recurrence patterns using radiomic analysis, suggesting the feasibility of using radiomics as imaging biomarkers to identify RT-resistant cases. This may facilitate early intervention for NPC patients^[19].

Currently, there is no set protocol to identify patients requiring adaptive radiation therapy (ART). During the course of intensity modulated radiotherapy (IMRT), oncologists may consider ART in some patients based on clinical factors such as significant weight loss, changes in palpable cervical lymph node volume, ill-fitting masks, or prolonged treatment breaks. NPC patients at the same clinical TNM stage may respond variably in terms of tumour size and lymph node volume after IMRT. In many cases, the decision of employing ART is dependent on the individual clinicians’ judgment and is not standardized. Lam et al. used CT-based imaging protocol for nodal radiomic analysis in predicting the use of ART, demonstrating the radiomics model outperformed the clinical staging model significantly (P < 0.0001)^[23]. The choice of CT was pragmatic and logically sound for its robust, shorter scanning time and good lymph node identification quality in ill-fitting thermoplastic casts (ifTMs). CT offers quick re-assessment of cases with extensive cervical lymphadenopathy before treatment and subsequent shrinkage of the disease during radiotherapy. In addition, studies were able to stratify high-risk NPC patients using the pre-treatment MRI-based radiomic analysis for ART eligibility assessment, facilitating oncologists in streamlining resources to enhance efficiency^[20,22].

RADIOMICS STUDIES IN TREATMENT OPTIONS PREDICTION

NPC with nodal metastasis at presentation and time of diagnosis is not uncommon. In fact, more than 75% of patients with NPC are diagnosed at advanced stages (i.e., Stages III and IV)^[5,24]. In addition to IMRT, chemotherapy and neoadjuvant chemotherapy are used as adjuncts in the treatment of locally advanced NPC. However, with the known acute and long-term toxicities from radiation and chemotherapy, not all patients are able to tolerate or respond well to the combined treatment modalities. Studies on deep learning and radiomics analyses attempted to predict the treatment response based on an individual’s clinical and biochemical status. By developing various radiomic-based models, therapeutic response prediction could make personalized treatment possible and avoid the unnecessary adverse side effects from chemoradiation. Xu et al. and Wang et al. developed MRI-based radiomic models on NPC lymph node metastasis cohorts to predict treatment responses in chemoradiotherapy (ChemoRT) and neoadjuvant chemotherapy (NAC), as shown in Table 3^[25,26]. The sensitivity and specificity of the radiomic models were higher in training cohorts than in validation cohorts. The AUCs of these models were over 0.9 for treatment response prediction for most studies.

RADIOMICS STUDIES IN SURVIVAL PREDICTION

The clinical staging of NPC is an important prognostic factor for survival. The one-year survival rates of NPC at stage I and stage IV diseases were 99.0% and 78.8% respectively. For stage I disease, the five-year survival has a small decline to 92.7%. However, the five-year survival of a stage IV NPC has a drastic drop to 47.1%^[24]. Despite a distinct geographical prevalence in southern China and Southeast Asia, over three-quarters of patients still present late with a neck mass due to its asymptomatic nature at an early stage^[24]. With the development in radiomics research, the prognostic performance of clinical staging is often compared with the novel development of radiomic models on survival analysis in locally advanced NPC. Studies focusing on the use of radiomics in predicting survival are listed in Table 4. Yang et al. developed a multidimensional nomogram using nodal volume-based radiomic features to predict progression-free survival, superseding the model using the clinical tumor-node-metastasis (TNM) staging^[27].

Similar to Yang’s study, all seven studies published from 2019 to 2021 suggested that radiomic models or combined clinical and radiomic models performed better than conventional clinical models in predicting distant metastasis and survival. There are only four studies that provided the quantity of radiomic features used, ranging from 208 to 2,364, for model development and training for reference^[27-30].

Locoregional recurrence free survival (LRFS) and distant metastasis free survival (DMFS) are of particularly concern to clinicians in the management of locally advanced NPC with lymph node metastasis. The variability of treatment options, such as neoadjuvant or induction chemotherapy, combined concurrent chemoradiotherapy, IMRT, or targeted therapy, are often compared with its efficacy, survival outcomes and toxicities in locally advanced NPC. Radiomics models could provide useful information in guiding treatment options and predicting survival outcomes based on the imaging-specific “tumor-fingerprint” of each individual patient.

DISCUSSION

This article reviewed studies specifically on lymph node metastasis in NPC for radiomics analysis to evaluate several objectives: risk stratification, treatment response, and survival predictions. Notably, the application of machine learning for the clinical management of NPC commenced more than a decade ago. Aussem et al. used a Bayesian network in machine learning to classify and identify risk factors in NPC in 2012^[34]. Ng et al. observed exponential growth in using AI in NPC since 2018^[7]. While there is a shift of interest from gross primary tumor volume to cervical lymphadenopathy in NPC for imaging analysis, studies of radiomics targeting lymph node metastasis in NPC have drawn researchers’ attention. Despite the growing number of radiomics studies on locally advanced NPC, they are often piecemeal and fall short of knowledge-building in the area of NPC management. This paper attempted to review and classify relevant pieces of literature according to clinical use. Hence, the knowledge could be better integrated for unleashing the future potential of radiomics in the treatment of NPC with lymph node metastasis.

As a clinician, the concepts of radiomics and deep learning in NPC appear to be a “black box” compared to traditional statistical models for prediction and survival analysis. The lack of understanding and explanation of how the model was built by image feature selection and algorithms raises clinicians’ skepticism about the generated outcomes. Critically analyzing the pitfalls of these studies are difficult from a clinician’s perspective, creating a huge resistance to translating these findings into real clinical applications. To improve the interpretability of AI models, Lundberg and Lee proposed a unified framework - Shapley additive explanations framework (SHAP) to understand how the machine learning models predict based on various weighting of selected features^[35]. For example, Chen et al. adopted SHAP to highlight the six features (BMI, EBVDNA, LDH, tumor volume, lymph node volume, and diameter) contributing to the complex nonlinear relationship with distant metastasis in NPC^[31]. SHAP was also used in nonmetastatic NPC scenarios. Du et al. suggested tumor shape sphericity, first-order mean absolute deviation, T stage, and overall stage influenced the outcome of 3-year disease^[36]. Although SHAP provides an explanation of the pattern learned by radiomic models, it does not correlate nor explain the patient’s characteristics with the imaging trained in the models. Yet, it does provide insights to clinicians in understanding the chain of thoughts of the machine learning models in predicting the outcomes in radiomics analysis.

Besides, whether the models reported could be extended from the training sample to the entire population is a key factor for promoting radiomics analysis into clinical use for the treatment of lymph node metastatic NPC. Most of the current studies involved retrospective data in a single or a few institutions. Since different imaging parameters affect the radiomic features being extracted, reproducibility of the data and feature selection become the major barrier. The major challenge in consideration of the clinical implementation of radiomics is whether the developed models are applicable across different institutions. Tumor segmentation is part of the key steps for radiomic implementation. The contouring of both primary tumor and nodal disease volume from imaging for radiomic features extraction is operator dependent and time-consuming even for an expert radiologist or oncologist. Likewise, the selection of radiomic features requires prior image processing and tumor contouring. Standardization and automation of this process may be a possible direction for this aspect. Automatic segmentation has been shown to reduce the time of radiotherapy planning by up to 30%-40%^[37]. The reduction of intra- and inter-observer variability during tumor delineation has also improved the consistency of IMRT performance^[37]. Different algorithms, such as single-atlas-based, multi-atlas-based, and deep-learning-based auto segmentation models were developed to delineate the head and neck lymph node metastasis in automatic tumor contouring. Atlas-based auto segmentation takes a collection of image segmentations as an atlas to map and delineate the region of interest of a new patient’s image data set^[38-40]. It could help to identify head and neck lymph node lesions^[39]. Although atlas-based auto segmentation could reduce the time for lymph node contouring and selection bias, there is no single best atlas to fit all individual image data^[38]. Multi atlas-based auto segmentation, using several patients’ tumor segmentations or peer-reviewed manual contours, provides more reference atlases to improve the accuracy in tumor contouring, at the expense of high computational power requirement^[40,41]. Deep learning based auto segmentation, most commonly with convolutional neural networks algorithm, provides robust and consistent lymph node segmentation tasks in head and neck regions^[41,42]. Using the automated approach for segmentation and radiomic feature extraction, it would pave the way for better radiomics model generalizability.

Factors that affect the radiomics variabilities due to image acquisition are collectively known as the batch effect^[43]. The protocol of medical imaging taken for NPC could be different among different centers. This heterogeneity predisposes the difficulty in the standardization of protocol development. In addition, the toolboxes for image pre-processing and radiomic feature definitions may lead to unwanted variations, affecting the generalizability of the model produced.

FUTURE DIRECTION

There is so far no consensus to interpret the NPC nodal metastasis within the imaging study - whether the number of lymph nodes, its size, or the radiological features of the malignant nodal lesion (e.g., the presence of extra-nodal extension) that would affect the radiomic model. The major future directions of radiomics studies are to address the issue of a batch effect in image acquisition and feature extraction standardization, and ultimately to improve radiomics reproducibility and generalizability^[14].

CONCLUSION

This review presented a summary of the vast applications of artificial intelligence in using radiomics and deep learning analysis on lymph node metastasis for advanced NPC. With the emerging technologies in the radiomic features applications, deep learning models have been readily developed and validated with good accuracy in auto-segmentation for tumor staging and treatment planning; risk-stratification for providing more personalized treatment options, and predicting the overall survival and disease-free survival.