Artificial intelligence to predict outcomes of head and neck radiotherapy

Highlights

• •Head and neck cancer radiotherapy lacks predictive tools due to its heterogeneity.
• •Deep learning models can successfully integrate clinical data and radiomics.
• •They promise to predict different outcomes to guide individualized decision.
• •Understanding, validating, expanding them is critical for clinical implementation.

Abstract

Head and neck radiotherapy induces important toxicity, and its efficacy and tolerance vary widely across patients. Advancements in radiotherapy delivery techniques, along with the increased quality and frequency of image guidance, offer a unique opportunity to individualize radiotherapy based on imaging biomarkers, with the aim of improving radiation efficacy while reducing its toxicity. Various artificial intelligence models integrating clinical data and radiomics have shown encouraging results for toxicity and cancer control outcomes prediction in head and neck cancer radiotherapy. Clinical implementation of these models could lead to individualized risk-based therapeutic decision making, but the reliability of the current studies is limited. Understanding, validating and expanding these models to larger multi-institutional data sets and testing them in the context of clinical trials is needed to ensure safe clinical implementation. This review summarizes the current state of the art of machine learning models for prediction of head and neck cancer radiotherapy outcomes.

Introduction

Global approximate incidence of head and neck cancer (HNC) is 880 000 patients each year worldwide [1]. As a core therapeutic option, radiotherapy (RT) is being used in 75 % of cases, combined with other treatment modalities such as chemotherapy or surgery [2]. Current HNC RT is associated with high rates of toxicity as well as adverse impacts on patients’ quality of life [3]. HNC encompasses a heterogeneous group of cancers originating from various subsites that are associated with various risk factors including viral infections [4], tobacco and alcohol use [5]. In addition to known heterogeneity, RT efficacy and tolerance also vary across apparently similar patients presenting with the same cancer subtype, anatomical stage and apparent risk factors [6]. Patient-specific clinical, radiological and biological factors are thought to drive these individual cancer outcomes. To date, HPV-positive status is known to be a favorable predictor of response to treatment in oropharyngeal cancer and have led to the concept of de-escalation treatments, but which can not be used outside of clinical trials as studies have been inconclusive [18]. Currently, there are no other biomarkers used to guide decisions in HNC RT.

Over the last decades, the refinement of RT techniques, along with the parallel developments in image guidance for better delineation and tumor localization over the course of treatment [7], [8], have been associated with an increasing interest for individualized RT with the aim of increasing or maintaining tumor control and reducing radiation toxicity [7]. Machine learning (ML) consists in the analysis of large amounts of empirical data using computational algorithms and leading to automatic extraction of discriminative features and learning of complex patterns [9], [10]. Developments in artificial intelligence (AI), particularly ML and deep learning (DL), have led to significant enthusiasm for the concept of “rapid learning health system”, whereby decision-making would be individualized based on analysis of large patient cohorts [11]. AI offers a unique opportunity for the development of predictive models that can help stratify individual patient’s risk and guide therapeutic decisions for optimal patients’ outcomes and quality of life in HNC RT. Herein, we review the current stance of the role of AI in predicting toxicity and therapeutic outcomes in HNC patients treated with RT.

Radiation oncology and artificial intelligence

The term “Artificial intelligence” first appeared at the Dartmouth Scientific Conference in 1956 [12]. ML is a subfield of AI that uses an algorithm to find patterns within data. DL is a subcategory of ML based on the use of neural networks and representation learning (Fig. 1).

Fig. 1.

Various levels of Artificial intelligence, including machine learning and deep learning.

Since the success of computer vision based analysis due to the increase in access to graphical process units (GPU), the field of medical image analysis has seen rapid benefits from ML approaches, which contributed in advancing the fast growing field of adaptive RT. With the increasing digitization and standardization of image acquisition and storage processes, the primary drawback of ML methods - the large training data requirements - continues to diminish. More specifically, recent studies focused on opportunities to use big data as decision support by predicting tumor response and toxicity outcomes [19], [20]. ML-based risk stratification could support decisions such as RT or systemic therapy intensification (or de-intensification), or even guide prophylactic measures in prevention of expected toxicity.

Radiomics (Fig. 2) is a research field which consists in the extraction of a large number of quantitative, hand-crafted features from different medical images (CT; MRI; PET-CT) - that cannot be manually deciphered by clinicians [21], and which will be then integrated to data-characterization algorithms [22]. Although pathology can be gold standard, it is not feasible to biopsy every node or every area of the tumor and radiomics offers the promise of characterizing different areas of the cancer [23]. Radiomics offer a particular potential in HNC RT as an immense volume of different imaging modalities are gathered routinely. Most HNC patients begin with initial diagnostic and staging images including diagnostic and/or planning computed tomography (CT) and magnetic resonance imaging (MRI). Positron emission tomography (PET) is also frequently part of routine diagnostic (or planning) imaging in HNC [24], [25]. During RT, daily CBCT (cone-beam computed tomography) serves for patient positioning further contributing to the quantity of available anatomic imaging [26]. More recently, the MR-Linac, which is available in an increasing number of institutions, has allowed for daily online MRI, therefore further improving image guidance [27]. After RT completion, follow-up typically includes CT, MRI or PET, based on different factors and guidelines [28].

Fig. 2.

Schematic illustration of radiomic pipeline.

Classification models are used to cluster data into different groups by approximating a mapping from a different set of inputs to discrete a set of outputs, has been mainly used as a predictive ML algorithm for HNC. The main algorithms are logistic regression (LR), decision tree, random forest (RF), support vector machine (SVM), k nearest neighbor (KNN), naives Bayes and artificial neural network. These traditional models have allowed better understanding of medical images, but with the major disadvantage of the time consuming step involving extracting and selecting important image features [35]. To address this issue, convolution neural network (CNN), a DL model that automatically learns, extracts and selects important features have shown promise in imaging classification tasks [36]. For any AI models, including classifiers, overfitting is a common problem when an algorithm is too focused on the training data set that it can not be successfully applied to a new data set. In order to verify if the model is not overfitting, a k-fold cross validation could be done by randomly partitioning a data set into k mutually exclusive subsets. Different metrics can be used in order to evaluate a model’s performance. These include sensitivity, specificity, accuracy, precision and recall. Another measurement could be done with the Area Under the ROC Curve (AUC) ranging from 0 to 1 (0 for a model’s predictive performance is 100 % wrong and 1, as its performance is 100 % correct). The calculation of the concordance index (C-index) could also be measured for predictive performance, on a scale between zero and one with 0.5 indicating no prognostic value.

Table 1.

Summary list of potential use of AI algorithms in different steps in HNC RT.

Steps	Clinical Use	Algorithms
1) Diagnosis/Initial evaluation	Clinical/Pathological/radiological data processing	SVM, LR, RF, DT, KNN, Bayesian, Linear Discriminant Analysis, DL and combination of ML and DL
2) Treatment Decision Making	Treatment Decision Aid	DSS, SDM
	Outcome Prediction	SVM, LR, Bayesian, neural network, decision trees and combination
	Toxicities Prediction	Bayes, LR, KNN, SVM, SVC, RF, XGBoost PLR, RFC, IBDM, CART, CNN
3) Simulation	Synthetic Imaging Generation	Fuzzy c-means clustering, CNNs and GANs
4) Treatment Planning	Auto-segmentation	DSC, CNN dd-Resnet
	Dosimetric Optimization	Unet, DenseNet, GANs
5) Treatment delivery	Image Guidance and Motion Management	CNN, Bayesian, SVM,
	CBCT quality improvement	CNN

Abbreviations: CBCT, cone-beam computed tomography; CNN, convolutional neural network; CART, Classification and Regression Tree; DL, deep learning; DT, Decision Tree; DSC, Dice Similarity Coefficient; DSS, clinical decision support systems; KNN, k nearest neighbor; LR, logistic regression;ML, machine learning; GANs, Generative Adversarial Networks; IBDM, image based data mining; RF, random forest; RFC,random forest classifier; SDM, shared decision making; SVM, support vector machine; SVC, support vector classifier.

Prediction of toxicity

Different Normal Tissue Complication Probability (NTCP) models have been proposed in order to help understand and estimate the risk of different toxicities. The common traditional models are the Lyman-Kutcher-Burnman models based on simplified characterization of the radiation-dose response for each anatomy [37]. For different toxicities, some studies have shown good validation [38]. In clinical practice, radiation oncologists routinely use the Quantitative Analyses of Normal Tissue Effects in the Clinic (QUANTEC) dose–response guidelines to determine tolerance dose to organs at risk [39]. However, the predictive power of currently models is largely limited by the lack of integration of patient-specific risk factors [40], [41] and omission of the inclusion of radiomic features despite studies having shown their relevance [38]. In order to account for these issues and update the current NTCP models, different ML methods have been studied [43]. Table 2 summarizes the current literature on AI predictive models for toxicities.

Table 2.

AI predictive models for toxicities.

Authors	Endpoints	Number of HNC cohort	Algorithms	Performance	Important features
Dean et al. (2016) [53]	Mucositis	351	PLR, SVC, RFC	RFC with AUC 0.71	Volumes of oral cavity receiving intermed - high dose
Dijk et al. (2016) [46]	Xerostomia and sticky saliva at 12 months	249	LASSO regularisation	AUC 0.77	IBMs
Pota et al. (2017) [95]	PG shrinkage and Xerostomia at 12 months	37	fuzzy classification	AUC 0.86 and 0.79	Final volume PG shrinkage
Dean et al. (2017) [49]	Dysphagia	173	PLR, SVC, RFC	RFC with AUC 0.71	PM receiving >1 Gy/fraction
Cheng et al. (2017) [52]	WL during RT and EOT	391	CART	AUC 0.773 and 0.821	Dose in OARs, oral intake, N stage, pain, nausea
Beasley et al. (2018) [47]	Trismus	86	Linear regression	Rs of −0.45	IBDM clusters in ipsilateral masseter
Gabry s et al. (2018) [44]	Xerostomia at different timelines	153	7 classifiers and 6 feature selectors	AUC 0.74–0.88	Dose gradient in PGs, PG volume, PG eccentricity
Jiang et al. (2018) [45]	Xerostomia 3 months post RT	427	RLR, LLR, RF	AUC 0.70	Dose pattern in PG/SMG
Reddy et al. (2019) [50]	Hospitalization Feeding tube WL	2121	RF, GB, LR	AUC 0.640-0.751	–
Wojcieszynski et al (2019) [51]	Grade > 3 toxicity (90and180days)	437	PLR, RF, XGBoost	C-statistic 0.65 and 0.63	PTV integral dose and integral dose out of PTV
Zhang et al. (2020) [54]	RTLI post RT in different timelines	242 NPC	RF	AUCs 0.830, 0.773 and 0.716	Features from medial temporal lobe
Humbert-Vidan et al. (2022) [55]	ORN	140	3D CNN, DenseNet 21	AUC 0.71	Clinical dosimetric distribution

Abbreviations: AUC, Area Under the ROC Curve; CART, Classification and Regression Tree; GB, Gradient boosting; IBMs, image biomarkers; LLR, Local linear forest; LR, logistic regression; PLR, Positive likelihood ratio; PG, parotid glands; PTV, Planning target volumes; RLR, Rigid logistic regression; RF, random forest; RFC, random forest classifier; SMG, Submandibular glands; SVC, support vector classifier; XGBoost, extreme gradient boosting; 3D, three dimensions.

As xerostomia is known to be a common major early- and long-term toxicity of HNC RT, different studies reported on radiomic features from salivary glands, namely the parotid glands (PGs) or submandibular glands (SMGs). Gabry et al [44] compared the predictive performance of seven classification algorithms, six feature selection methods, and ten data cleaning/class balancing techniques by integrating retrospective dosiomic, radiomic and demographic data from 153 HNC patients treated with definitive RT. Important features such as dosimetric shape and gradients, PG volume and eccentricity, and DVH spread were identified. Their MLs based models performed better compared to the traditional NTCP models based on mean dose to PG only. Jiang et al [45] tested three different algorithms applied to a data set including clinical and radiomic features such as voxels dose in the PGs and SMGs from 427 HNC patients treated with definitive RT. Similar to. Gabry et al, the authors were able to observe that specific dose patterns across the subvolumes in both organs were an important predictive feature and their ridge logistic regression model had the best performance with AUC of 0.70. Dijk et al [46] used the CT based image biomarkers (IBMs) of the PGs and SMGs from 249 HNC patients treated with definitive RT in order to improve the NTCP predictive models for sticky saliva (STIC) and moderate-severe xerostomia 12 months. For both toxicities, pre-selection through a lasso regularization identified different important radiomic features and the predictive performance of NTCP models were stronger when IBMs were added (AUC 0.74 vs 0.77). Beasley et al [47] also used ML applied to the image based data mining (IBDM) in order to identify clusters of dose distribution voxels involved in radiation induced trismus. From clinical and radiomic data set from 86 HNC patients focusing on the dose distributions within the anatomy of interest, different features investigated in a multivariable analysis and an internal/external validation demonstrated the importance of dose patterns within anatomy to predict trismus.

Dysphagia is also a common toxicity among patients treated with RT for HNC causing a major impact in patients’ quality of life [48]. In order to improve the current predictive NTCP model for dysphagia, Dean et al [49] incorporated spatial dose metrics in different ML models with a prospective data set of 173 HNC patients including clinical and dosimetric features focusing on pharyngeal mucosa (PM). Their RFC model with a highest AUC of 0.71 identified that the volume, length and circumference of PM receiving 1 Gy/fraction and higher were strongly associated with the risk of dysphagia. More recently, Reddy et al [50] used a different data set from 2,121 HNC patients in order to compare predictive performance of three different classifiers for unplanned hospitalizations, feeding tube placement and significant weight loss. This method identified over 700 treatment-related and clinical variables, and achieved AUC values of up to 0.64, 0.75, and 0.75 for RF, gradient boosting, and LR, respectively. Wojcieszynski et al [51] compared the predictive performance of three ML methods on a prospective data set of 437 HNC patients treated with definitive chemoradiation (CRT). Their RF model yielded moderate success for toxicity at 90 and 180 days with c-static of 0.65 and 0.63, respectively. From this study, higher integral doses outside of the target volume, target volume integral dose, body mass index (BMI) and age were important factors associated with increased grade 3 + toxicity. Cheng et al [52] used a Classification and Regression Tree model in HNC patients treated with definitive RT by using demographic, dosimetric and clinical data from 391 patients in order to predict weight loss ≥5 kg at 3 months post-RT. Two models were built, one during the RT planning and one at the end of the treatments. When additional treatment-related data was added to each model, the predictive performance was improved, with an AUC of 0.77 and 0.82, respectively.

Several other radiation-induced toxicities have been the focus of individual studies. Mucositis was the focus of a study by Dean et al [53], where clinical, dose-volume and spatial dose metrics data from 317 HNC patients were used to build predictive ML models for severe acute mucositis. Among different models tested, the discriminative performance was not improved with the additional spatial dose metrics. Important features were in the range V80-V220 and the most important feature was the V220. Interestingly, in contrast to the RTOG guidelines focusing on mean dose to the oral cavity, the authors therefore identified that the strongest feature associated with severe acute mucositis was the volume receiving intermediate and high doses. To predict radiation-induced temporal lobe injury, Zhang et al [54] used retrospective clinical and CT and MRI based radiomic data from 242 nasopharynx (NPC) patients treated with definitive RT. Different radiomic features were first extracted from the medial temporal lobe regions. RF predictive model showed strong predictive performance in three subsequent radiological follow-ups preceding the onset of radiation-induced temporal lobe injury with the mean AUCs of between 0.71 and 0.83. More recently, Humbert-Vidan et al [55] used retrospective demographic, clinical and dosimetric data (3D dose distribution map) of 140 HNC patients (70 patients with ORN and 70 patients as control group) in order to compare the predictive performance of a 3D densely-connected 121-layer convolutional neural network (CNN) model with a DVH based RF model. The 3D DenseNet121 CNN model had better performance with an average AUC of 0.71 (0.64–0.79), compared to 0.65 (0.57–0.73) for the RF model.

There was a general trend of increasing feature importance with increasing dose and feature importance was also high for RT dose metrics in the range V80 – V220.

Prediction of cancer control outcomes

Current predictive models for HNC are mainly based on the TNM (Primary tumor, regional lymph nodes and distant metastasis) staging [56], which guides oncologists in the selection of the appropriate therapeutic options for patients. In oropharyngeal cancer (OPC), other prognostic variables established in the context of clinical trials, namely human papillomavirus (HPV) status and tobacco smoking pack-years, have largely been integrated in the clinical practice. While patients with HPV-associated OPC generally have an improved prognosis, we now recognize that a subset of these patients present a highly aggressive behavior [57]. There is therefore a general concern that safe treatment de-escalation should not jeopardize the chance of cure of these patients, and that more reliable tools to predict tumor behavior are greatly needed. In recent years, several predictive AI models to better predict cancer control outcomes in HNC have been published and are presented in Table 3.

Table 3.

Different AI predictive models for outcomes.

Authors	Patient population	Predicted outcomes	Number of patients	Algorithms	Performance	Important features
Parmar et al. (2015) [63]	HNC SCC	3Y OS	206	13 feature selectors and 11 ML classifiers	3 feature selectors (MRMR, MIFS and CIFE) with AUC 0.66–0.69 3 classifier (RF, NN and BY) with AUC 0.61–0.67	CT based radiomic features
Jiang et al. (2015) [62]	NSC	OS	347	SVM	HR 3.45	Combination of CRT
Li et al. (2018) [65]	NSC	LR	306	PCA, ANN, KNN SVM	ANN with accuracy of 0.812	MRI based radiomic features
Zdilar et al. (2018) [64]	OPC	OS RFC	529	MRMR, Wilcoxon, RF, RReliefF, RRF, IAMB, RSF, PCA	RF selectors AUC 0.75 and C-index 0.68	CT based radiomic features
Fujima et al. (2019) [96]	Sinonasal SCC	LF	36	SVM	Accuracy of 0.96	MRI based radiomic features
Wu et al. (2019) [97]	OPC	DMFS	140	RSF	C-index 0.73	Max distance between nodes and tumor-nodes
Zhou et al. (2019) [67]	NPC	DM	176	PyRadiomics features extraction, features selections (Mann-Whitney U test, mRMR, Lasso), LR	AUC 0.827 (training group) and 0.792 (validation group)	MRI based radiomic features
Tseng et al. (2020) [58]	Oral cavity	Survival (Cancer specific and loco-regional recurrence free)	334	Elastic Net Penalized Cox Proportional Hazards regression	C index 0.689 and 0.693. Distant metastasis free survival not different	Genetic data
Howard et al. (2020) [59]	HNC SCC	OS	33 527	DeepSurv, RSF N-MTLR	HR of 0.79, 0.83 and 0.90	Stage T4, HPV status, tonsil subsite
De Felice et al. (2020) [32]	OPC	OS	273	RF - Classification tree	Mean decrease accuracy of 4.29, 2.49 and 1.11 %	HPV status N status Early responders
Tran et al. (2020) [68]	HNC SCC	Local nodal response	32	LR, KNN, naive-Bayes	Accuracy 87.5 % with three feature model	Quantitative US radiomic features
Tosado et al. (2020) [70]	OPC	OS, RFS	644	RReliefF feature selector, Cox Model, RSF	AUC 0.6395 (OS) and 0.6483 (RFS)	Radiomic features combined with clinical features
Bogowicz et al. (2020) [71]	OPC, hypo pharynx, larynx, oral cavity	2YOS, HPV	1174	Feature selector (LR Z-Rad), Classification (hierarchical clustering, LR)	No significant differences in AUC between centralized and distributed.	981 radomic features
Rich et al. (2021) [69]	OPC HPV +	DM	225	Feature extractor (SMOTE, ADASYN, borderline SMOTE), SVM	AUC 0.84–0.95	CT based radiomic
Le et al. (2022) [72]	HNC SCC	10Y DM, Lr, OS	371	Cox Model, RF, CNN, DenseNet, InceptionV3, ResNet, ResNeXt and PreSANet	PreSAnet - Accuracy of 74 % (Lr) and 79 % (OS)	Performance decrease with PET images

Abbreviations: ADASYN, adaptive synthetic sampling; ANN, artificial neural network; BY, Bayesian; CT, Computed tomography; C-index, concordance index; CIFE, conditional informax feature extraction; CRT, chemo-radiotherapyDM, distant metastasis; DMFS, distant metastasis-free survival; HNC, head and neck cancer; HPV, human papillomavirus; HR, Hazard ratio; IAMB, incremental association Markov blanket; KNN, k nearest neighbor; LF, local failure; Lr, locoregional recurrence; LR, logistic regression; LRC, loco-regional control; NPC, nasopharyngeal cancer; MIFS, mutual information based feature selection; MRI, Magnetic resonance imaging; MRMR, Minimum redundancy feature selection; N, node; NN, nearest neighoor; N-MLTR, Neural Multi-Task Logistic Regression; OS, overall survival; PET, positron emission tomography; PCA, Principal component analysis; RF, random forest; RFC, relapse-free survival; RRF, Regularized random forest; RSF, random survival forest; SCC, squamous cell carcinoma; SMOTE, synthetic minority over-sampling technique; SVM, support vector machine; US, Ultrasounds; 2D CNN, two dimensions convolutional neural network; 3D CNN, three dimensions convolutional neural network.

To better stratify survival outcomes, Tseng et al [58] evaluated the elastic net penalized Cox proportional hazards regression-based risk stratification model in operated oral cavity cancers. The authors integrated clinicopathologic and genomic data from 334 patients with locally advanced HNC treated with curative intent surgery, combined with adjuvant RT or CRT. Compared to the baseline model using clinicopathologic data alone, the identification and integration of genetic features associated with prognostic led to a model with better classification performance, with mean C-indexes of 0.689 (vs 0.673) and 0.693 (vs 0.678) for cancer-specific survival and locoregional recurrence-free survival, respectively. Also in the post-operative setting, Howard et al [59] aimed to build an overall survival (OS) predictive model to better identify patients that may benefit from the addition of adjuvant concurrent chemotherapy. The authors evaluated different ML models integrating a large retrospective data from the National Cancer Database, including 33 527 patients with squamous cell carcinoma of the oral cavity, oropharynx, hypopharynx, or larynx treated with definitive surgery followed by an adjuvant RT or CRT. Three different models were trained, then validated using a subset of the same cohort. Compared to RTOG 95–01 and EORTC 22931 recommendations [60], [61], treatments guided by all three models had improved survival outcomes with hazard ratio (HR) of 0.79–0.90 and c index of 0.691–0.695 with similar accuracy and there was no survival benefit for CRT for patients recommended to receive RT alone. These models also identified important variables related to prognosis such as year of diagnosis, T4, HPV positivity and tonsillar subsite.

In order to evaluate the role of radiomics for risk stratification of DM, Rich et al [69] used CT based radiomic and clinical features from 225 locally advanced OPC HPV+ patients treated with curative intent RT or CRT. Nine different algorithms were built using different radiomic datasets derived using different algorithms integrated to SVM classifier. All classifiers achieved at least an excellent level in discriminating the two patient cohorts with DM or not. In order to build a predictive model for OS and RFS outcomes, Tosado et al [70] incorporated retrospective CT based radiomic data with clinical data from 644 OPC patients treated with RT or CRT. Distinctive radiomic and clinical features with the cluster labels were identified and different supervised models were built using these features. Compared to the baseline model based on clinical features only, these models had better predictive performance for both outcomes. In another study by Jiang et al [62], aiming to predict outcomes in metastatic NPC patients treated with chemotherapy alone, RT alone or the combination of both, the retrospective hematological, clinical and therapeutic parameters of 347 patients were used in a SVM model. The multivariate model had a strong performance with an AUC at 0.761 and the classifier was able to stratify the patients into low risk and high risk groups with significantly different 2 year OS of 71.4 % vs 18.8 %, respectively. This classifier also helped identify that combined CRT was associated with significantly better outcomes in the low risk groups, but not in the higher risk groups.

Other studies focused on integrating radiomic data into ML algorithms to optimize predictive models for HNC treatment outcomes. In order to identify the optimal ML methods for radiomic-based overall survival prediction, Parmar et al [63] compared the performance of 13 features selection methods and 11 ML classification methods integrating CT based 440 radiomic features from 196 HNC patients. Three feature selection methods had the best performance with AUC between 0.66 and 0.69 and stability between 0.66 and 0.70 compared to the median values of AUC = 0.61 and stability = 0.66. Three classifying methods had the best performance with AUC between 0.61 and 0.67. Zdilar et al [64] used retrospective CT based radiomic, clinical, demographic, toxicity and cancer control outcome data from 529 OPC patients treated with curative intent RT or CRT to compare the predictive performance of different selectors for OS and RFS. Among 3800 radiomic features extracted, selected features using 8 different methods resulted in better AUC compared to clinical features alone. Among the feature selectors, RF based selectors had the best overall scores. In order to build a model predicting a radioresistance, Li et al [65] used retrospective data from 306 NPC patients treated with definitive CRT. Clinicopathological and radiomic/dosimetric features from planning CT and from follow up imagings including CT, MRI or PET. Once detected, recurrent tumor volumes were delineated, then was categorized as “in field recurrence” if the recurrence was inside the high-dose target. Eight discriminative features were identified from pretreatment MRIs compared between the patients with and without the disease recurrence. Features were fed to three different MLs, which were trained, then validated and yield accuracies ranging between 0.732 and 0.812. These results could indicate possible differences in heterogeneity in LR tumors. For a predictive performance of loco-regional recurrence (LRR), Starke et al [66] used the retrospective data from 291 patients with locally advanced HNC treated with CRT. A baseline Cox proportional hazards model (CPHM) using clinical features alone was compared to different 3D-CNN and 2D-CNN models built from scratch combining clinical features and CT images. Among these, the ensemble of 3D-CNNs had the best performance and successful validation with a C-index of 0.31. Patient risk group defined by this model's predictions showed significant differences in LRR with p = 0.001. The C-index for 2D-CNN and for CPHM was 0.38 and 0.39, respectively. In order to satisfy risks for DM and 5Y OS, Zhou et al [67] used MRI based radiomic with dosiomic and clinical features from 176 NPC patients treated with curative intent RT or CRT. With the radiomic features extracted, an algorithm model was built in order to classify into high- and low risk groups for DM. With the clinical features, the radiomic based models show strong predictive performance for both in the training and validation cohorts with AUC of 0.827 and 0.792 respectively. Another study was done to predict therapeutic response in metastatic lymph nodes by Tran et al [68]. Their team used the quantitative ultrasound based radiomic markers from 32 HNC patients with positive lymph nodes treated with curative intent RT or CRT. Depending on their 3 months follow up MRI, patients were divided into two different categories: complete responders or partial responders. Different radiomic features were extracted, then applied in LR, KNN and a naive-Bayes. Multi Parametric models showed a strong predictive power with high accuracy of 87.5 %. Significant differences in radiomic parameters were found between the two groups.

In order to evaluate the role of radiomics for risk stratification of DM, Rich et al [69] used CT based radiomic and clinical features from 225 locally advanced OPC HPV+ patients treated with curative intent RT or CRT. Nine different algorithms were built using different radiomic datasets derived using different algorithms integrated to SVM classifier. All classifiers achieved at least an excellent level in discriminating the two patient cohorts with DM or not. In order to build a predictive model for OS and RFS outcomes, Tosado et al [70] incorporated retrospective CT based radiomic data with clinical data from 644 OPC patients treated with RT or CRT. Distinctive radiomic and clinical features with the cluster labels were identified and different supervised models were built using these features. Compared to the baseline model based on clinical features only, these models had better predictive performance for both outcomes. In another study by Jiang et al [62], aiming to predict outcomes in metastatic NPC patients treated with chemotherapy alone, RT alone or the combination of both, the retrospective hematological, clinical and therapeutic parameters of 347 patients were used in a SVM model. The multivariate model had a strong performance with an AUC at 0.761 and the classifier was able to stratify the patients into low risk and high risk groups with significantly different 2 year OS of 71.4 % vs 18.8 %, respectively. This classifier also helped identify that combined CRT was associated with significantly better outcomes in the low risk groups, but not in the higher risk groups.

The sufficient quantity of a radomic data set is a common problem for an AI algorithm development and multicenter approach can be a solution but implicating ethical issues. In order to address this issue, Bogowicz et al [71] tested the distributed learning technique enabling training models on multicenter data without data leaving the hospitals. Two different approaches, centralized and distributed, were compared for 2Y OS and HPV status predictive models built with CT based radiomic, dosiomic and clinical features from 1174 HNC patients treated with curative intent RT or CRT. For both feature selection and classification, there was no significant difference in terms of performance between these two approaches. Most recently, Le et al [72] used retrospective cross-institutional patho-clinical and PET-CT based radiomic data set from 298 HNC patients treated with curative intent RT or CRT in order to train a predictive model based on a pseudo-volumetric convolutional neural network with PreSANet. The model was internally validated, then an extensive set of ablation experiments on the public data set showed AUROC of DM, LR and OS between 80 and 82 %. External validation on a retrospective dataset showed an AUROC at 69 % and a validation of single site-holdout and cross-validation showed mean accuracy across four different institutions was between 70 and 72 %.

Discussion

Recent studies have shown promising results in the use of ML in the field of HNC RT in predicting therapeutic outcomes and toxicity. Different algorithms have shown good predictive performance and have helped identify features that provide insight into the heterogeneous nature of HNC. Those features have included demographic characteristics, molecular, dosimetric, radiomic and therapeutic factors. Their integration to the current clinical decision algorithms has the potential to improve risk stratification and selection of optimal therapeutic options.

Despite promising results, these models remain largely premature for clinical use at this stage. One major concern is the lack of standardization of the largely retrospective data used in a single center. In radiomics specifically, for each imaging modality, intra and inter-institutional variations in scanner, acquisition and reconstruction parameters have been shown to impact the robustness of the predictive models [73]. Pertaining to RT data, differences in tumor and organs at risk segmentations as well as dosimetric data further challenges reproducibility [42], [74]. In addition, variations in feature extraction techniques, choice of robustness metrics and outcome definitions used across studies complicate the interpretation of results. The small sample sizes and high heterogeneity of HNC and frequent lack of external validation lead lack of generalizability of the current studies. Increased use of open-access data sets and multicenter prospective cohorts, along with strict guidelines for data standardization would be critical for these models to reach clinical usability [71], [72], [75]. In addition, training clinical oncologists to the field of AI, as well as integrating them early in the development and validation of these models will be increasingly important. As clinicians are ultimately responsible for decision making, ensuring their adequate understanding and interpretation of these models will help overcome the ‘’black box’’ problem and facilitate clinical implementation of these tools [76], [77].

Beside the risk stratification and decision support in HNC RT, AI has substantial potential in other upstream tasks such as automatic detection and segmentation of anatomical structures [13], automatic registration of mono-modal or multimodal images [14], temporal motion compensation [15], tumor and lesion grade classification [16], and nomograms for risk stratification-prognostic modeling [17]. AI can independently be introduced at all stages of a patient’s treatment from diagnosis, to planning and re-planning, to long-term follow-up and prognosis, while allowing or benefiting from expert input along the way.

AI can reveal the radiomic signatures in HNC such as tumor characteristics including HPV status [29] or Programmed Death-Ligand 1 expression [30], identification of extranodal extension [31] as well as cancer control outcomes [32] and larynx/hypopharynx cancers [33]. Radiomics also has the potential to provide a quantitative assessment of tumor and normal tissue reaction to RT over the course of treatment (i.e. delta-radiomics) [34]. The emerging use of multi-omics ((Gen-omics, Epigen-omics, Transcipt-omics, prote-omics, metabol-omics and microbi-omics) [78], [79] in oncology, along with the increasing quantity and quality of imaging in radiation oncology, represent a clear opportunity to propule the role of AI in HNC. The integration of daily imaging used over the course of RT in order to capture dynamically tumor response or early signs of toxicity could guide therapeutic decision or early interventions more efficiently, as suggested in early work using CBCTs data in dynamic predictive algorithms [80], [81]. The increasing availability of the MR-Linac technology across institutions will lead to increased quality and both anatomic and functional information of daily RT imaging and will further increase the potential to unlock dynamic imaging-based biomarkers [82], [83], [84], [85]. Finally, as the field of liquid biopsy if expanding rapidly, non invasive biologic serum or salivary HNC biomarkers [86], [87], [88] could be further integrated to these algorithms to increase the precision of dynamic clinical outcomes predictions over the course of treatment and in post-treatment follow-ups [89], [90], [91], [92], [93], [94].

This review has several limitations. First, its non-systematic review method could cause potential risks of different bias with less transparency. However, this method seemed suitable to offer a general overview of the current stance of AI in HNC RT. Also, our study did not include a recently developed radiomics quality score, a tool built by Lambin et al. in order to determine the validity and completeness of radiomics studies [98]. For a more comprehensive understanding of the subject matter, future reviews could include use of this tool as well as an insight on the ongoing clinical trials evaluating AI and radiomics tools in HNC RT [99], [100], [101].

AI-models predicting cancel control and toxicity outcomes in HNC RT have shown promising performance and would be of high clinical utility for individualized risk-based decision making. These important challenges to the development and safe clinical implementation of these models could only be overcome with coordinated collaborative efforts to standardize, validate and expand these models to large enough datasets and test in the context of clinical trials.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.