Efficacy and Quality-of-Life Following Involved Nodal Radiotherapy for Head and Neck Squamous Cell Carcinoma: The INRT-AIR Phase II Clinical Trial

This article is featured in Selected Articles from This Issue, p. 3251

The radiotherapeutic management of head and neck squamous cell carcinoma (HNSCC) has historically involved the irradiation of both gross disease as well as “elective” volumes encompassing the primary tumor and neck, aimed toward sterilizing clinically and radiologically occult disease with lower doses of radiotherapy. Although the standard-of-care radiotherapy technique has transitioned to intensity-modulated radiotherapy (IMRT), the basic treatment paradigm of delivering 50 to 60 Gray (Gy) of elective neck irradiation (ENI) to normal-appearing cervical nodal basins has remained the same. Because salivary and swallowing toxicities are related to the dose to the elective neck, there have been several prospective trials investigating elective neck dose deescalation, reducing the dose to 30 to 40 Gy—without increased failure in elective regions (1–3).

However, even low radiation doses to the cervical neck can increase toxicity to normal organs-at-risk (OAR) such as salivary glands and pharyngeal constrictors, and there are other poorly characterized risks of large-volume neck irradiation, such as carotid artery damage and soft-tissue fibrosis (4, 5). In several other disease sites, involved nodal radiotherapy (INRT) has become the standard-of-care, such as lymphoma and lung cancer (6), leading to a reduction in toxicity without compromising oncologic outcomes. However, such a paradigm has not been meaningfully explored in HNSCC due to the perceived high risk of occult disease and subsequent regional recurrence.

Whereas ENI can sterilize micrometastatic nodal disease without explicitly identifying individual lymph nodes (LN) that may harbor cancer, INRT requires targeting each individual LN that may contain malignancy. Some LNs are clearly positive because of their large size or high activity on PET-CT, but there is often significant uncertainty about the malignant potential of smaller and less 2[18F]fluoro-2-deoxy-D-glucose (FDG)–avid LNs in HNSCC (7). To facilitate the treatment of LNs harboring occult disease, we developed an artificial intelligence (AI)–based radiomics model to classify the malignancy status of individual LNs by integrating both handcrafted and deep-learning–extracted imaging features from CT and PET (8–10). In this prospective phase II study entitled Involved Nodal Radiotherapy Using AI-Based Radiomics (INRT-AIR), ENI was completely eliminated for patients with oropharyngeal, laryngeal, and hypopharyngeal squamous cell carcinoma (SCC) treated with radiotherapy or chemoradiotherapy (CRT). This INRT approach presents a new treatment paradigm for HNSCC, and this INRT-AIR study is the first prospective clinical trial to investigate INRT for HNSCC treatment under the guidance of AI.

This prospective phase II study was approved by the University of Texas Southwestern (UTSW, Dallas, TX) Institutional Review Board. Patients were enrolled from August 2019 through September 2020 and included patients seen at both UTSW Medical Center and Parkland Hospital. All patients signed informed consent prior to enrollment on the study, which was consistent with the principles in the Declaration of Helsinki.

Patients with newly diagnosed SCC of the oropharynx, larynx, or hypopharynx who were candidates for primary radiotherapy treatment were eligible for enrollment. All stages of cancer were included, except glottic larynx T1–2N0 disease. Patients were required to have a neck CT or MRI and ¹⁸F FDG PET-CT within 60 days of registration. Patients were excluded if they had previous treatment of their head and neck cancer, had prior invasive malignancy with expected disease-free survival of less than 3 years, or were immunocompromised.

Patients with stage IVA-B and stage III (T3-N0–1; American Joint Commission on Cancer Stating; AJCC, 7) disease were treated with daily radiotherapy and concurrent chemotherapy. The chemotherapy regimens were at the discretion of medical oncology and included cisplatin, carboplatin-paclitaxel, or cetuximab. Patients with T1–2N1 disease were treated with either concurrent chemoradiotherapy or accelerated radiotherapy, per shared decision-making between the oncology team and patient. Individuals with stage II cancer were treated with either accelerated radiotherapy or daily radiotherapy alone, and stage I disease was treated with daily radiotherapy alone.

The gross tumor volume of the primary tumor (GTVp) was defined by clinical examination, CT, and PET-CT. The primary clinical target volume (CTVp) was a 5 to 8 mm expansion around the GTVp, excluding air and natural barriers to spread, such as the mandible or vertebral bodies. Involved nodes (GTVn) were defined by enlarged size (greater than 1 cm short-axis, 1.5 cm in level II), internal necrosis/ring enhancement, or standardized uptake value maximum (SUV_max) greater than 3.0.

There was no ENI. Instead, individual noninvolved LNs were identified as suspicious or potentially suspicious and treated on the basis of a protocol-defined algorithm. LNs were defined as suspicious if the axial cross-sectional summed diameter was 17 mm or more in any dimension, the FDG uptake of the node was greater than the internal jugular vein blood pool uptake (visual analysis; refs. 11, 12), the node was rounded or had irregular borders, any sized retropharyngeal node was seen, or a level VI node was identified in a hypopharyngeal primary.

A LN was defined as potentially suspicious if either it was in the same station as an involved node but did not meet any other criteria or it exhibited heterogeneous enhancement and was in the same or adjacent station as a suspicious or involved node. The LN needed to be present on at least two slices to be considered potentially suspicious. The treatment of potentially suspicious nodes was at the physician's discretion, but in practice, they were almost always included. Figure 1 summarizes this method and how it differs from conventionally designed radiotherapy.

In addition, all noninvolved nodes traversing at least two axial slices (3 mm in thickness) and not meeting any of these criteria were contoured separately and submitted to the AI-based classification module, and any node assessed with at least a 50% probability of malignancy on either the combined PET-CT or disease-specific CT model was labeled suspicious and treated. All AI-defined LNs were treated except level IB nodes in an otherwise LN-negative neck, which were at the discretion of the radiation oncologist. Therefore, all AI-identified nodes would not have been treated without the use of the model. The methodology used in the generation of the AI model is described in Supplementary Methods.

The planning target volume (PTV) for each GTV and the CTVp was created by a 5 mm universal expansion. The PTV expansion for suspicious and potentially suspicious nodes may be reduced to 1 mm if adjacent to an OAR. The PTV for the primary and involved nodes received 70 Gy in 35 fractions. The PTV for the CTVp received 63 Gy in 35 fractions. The PTV for the suspicious and potentially suspicious nodes was treated to 66.5 Gy in 35 fractions. Figure 2 shows an example dose distribution from a patient with oropharyngeal cancer.

Patients were all planned in the Eclipse treatment planning system and treated using volumetric modulated arc therapy (VMAT) on a Varian Truebeam or Vitalbeam linear accelerator; adaptive radiotherapy was not used in this study.

Patients underwent restaging PET-CT 11 to 14 weeks from the completion of treatment, and surveillance neck CT was performed at 6, 12, 18, 24, and 36 months following therapy. Patients completed four patient-reported outcome (PRO) questionnaires at baseline and in follow-up: European Organization for the Treatment of Cancer (EORTC) QLQ-30 and HN35, MDADI (MD Anderson Dysphagia Index), and EQ-5D.

To provide a comparison with the incidental dose to the standard elective neck volumes, the neck CTV for each patient was contoured and expanded by 5 mm to make the PTV per institutional standards. For oropharynx patients, the node-negative neck included levels II–IV, and the node-positive neck included levels IB–V and the retropharyngeal nodes. The neck CTV for patients with larynx cancer was similar except the retropharyngeal nodes were not covered, although the patients with hypopharynx cancer had the retropharyngeal nodes and level V included regardless of neck status. The V50 and V30 values for each elective neck PTV were recorded. The V50 was chosen since 50 Gy is the standard-of-care elective neck dose (13), and V30 reflects the dose (30 Gy) reported in a recent retrospective analysis of elective dose deescalation in oropharynx cancer (2).

The primary endpoint of the study was the 2-year crude risk of solitary elective volume recurrence, which focuses on first recurrences that would have been prevented with conventional ENI. The primary hypothesis was that the risk of elective or out-of-field neck failure preceding in-field or metastatic progression was less than 13% (i.e., 10% more than the baseline risk of 3%). The 10% threshold is a general rule to motivate irradiation of a given region, although surgical recommendations have raised the elective treatment threshold to 15%–25% (14). A sample size of 60 achieved a 90% power to detect a noninferiority difference of 0.1 using a one-sided exact test with a significance level (α) of 0.05, with a target accrual of 67 to account for 10% dropout.

Secondary endpoints of the study included PRO assessment, acute, and late swallowing toxicity and dermatitis, as well as overall survival (OS), progression-free survival (PFS), and the cumulative incidences of local, regional and distance recurrence. Survival analyses were performed using the Kaplan–Meier method, and cumulative incidence rates were calculated with death as a competing risk. Changes in PRO comparisons were analyzed with a paired t test. Statistical analyses were performed with SAS 9.4.

All analyzed clinical trial data presented in the article, including the study protocol, are available upon reasonable request by contacting the corresponding author.

Baseline patient and disease characteristics are described in Table 1. The median age of patients was 62 years [interquartile range (IQR), 55–70], with the majority of individuals white and male. Oropharyngeal cancer was the most common diagnosis (n = 48, 72%), although only 27 patients (40% of the entire cohort) were p16+, T1–3 N0-N2b (AJCC 7) oropharyngeal SCC with no more than 10 pack-years smoking history (defined as favorable oropharyngeal cancer). Definitive chemoradiotherapy (typically with cisplatin) was the most frequently delivered treatment (91%).

Table 2 identifies the nodal levels treated using this paradigm as well as the frequency in each station that the manual radiologic criteria or AI model identified a suspicious node. The AI model identified a nontrivial percentage of potentially involved nodes, most commonly in levels II–IV. There were a mean 31.1 (SD, 18.4) and median 30 (IQR, 18.5–41.5) nodes per patient submitted to the AI module, of which a mean 3 (SD, 2.3) and median 2 (IQR, 1–4.5) nodes were classified as positive. The mean OAR metrics for oropharynx and larynx/hypopharynx patients are displayed in Table 3.

The dose–volume histograms to the hypothetical elective neck PTV revealed a mean (SD) V50 and V30 of 41.6% (10.4%) and 58.4% (13%), respectively, and a median (IQR) V50 and V30 of 40.7% (33.1%–49%) and 57.7% (50.3%–68.3%), respectively.

At a median follow-up of 33.4 months from enrollment (IQR, 28.2–37.4 months) and 31.6 months (IQR, 26.6–35.6) from the end of radiotherapy for surviving patients, the 2-year OS and PFS probabilities were 91% and 82%, respectively. The 2-year OS for patients with oropharynx and larynx/hypopharynx cancer were 96% and 79%, respectively, and the 2-year PFS were 90% and 63% for oropharynx and larynx/hypopharynx cancer, respectively. For the 27 patients with favorable oropharyngeal cancer, the 2-year OS and PFS probabilities were 100% and 92%, respectively.

The cumulative incidences of locoregional recurrence (LRR), local recurrence (LR), regional recurrence (RR; Supplementary Fig. S1) and distant metastasis (DM) in the whole cohort were 11%, 9%, 3%, and 6%, respectively. By definition, LR means disease recurrence in the primary site, and RR is a nodal failure anywhere in the neck. Patients who experienced either an LR or RR (or both), were also coded as an LRR. Disease recurrence below the clavicles was considered a DM. The comparable LRR, LR, RR, and DM outcomes for patients with oropharyngeal cancer were 6%, 4%, 2%, and 9%, respectively, and 21%, 21%, 5%, and 0%, respectively, for patients with laryngeal/hypopharyngeal cancer. For patients with favorable oropharyngeal cancer, the cumulative incidences of LRR, LR, RR, and DM were 4%, 4%, 0%, and 4%.

There were two nodal failures, one of which developed in a previously involved LN and occurred simultaneously with an in-field primary recurrence. The other was an out-of-field elective failure with concurrent distant metastasis. Thus, the two-year risk of solitary elective nodal failure was 0%.

The one out-of-field nodal recurrence presented synchronously in the skull base, supraclavicular fossa, and with a metastasis in an AP window node approximately 13 months after completing CRT. The patient originally presented with an HPV-positive T4 N3 oropharyngeal cancer with an extensive marijuana history, which he continued to smoke after treatment. The nodal disease at presentation was intimately associated with the jugular sheath. He developed synchronous regional and distant metastases 13 months after completing CRT. He was treated with surgical excision of the supraclavicular node followed by adjuvant radiotherapy delivered concurrently with definitive radiotherapy to the skull base. He received stereotactic ablative radiotherapy to the AP window node. This salvage treatment rendered him NED (no evidence of disease), but he died approximately 5 months after the recurrent diagnosis, without a known cause of death. Retrospectively, a small supraclavicular node was visible on a pre-treatment neck CT (but not the simulation CT); the pattern-of-spread appeared contiguous superior and inferior to the jugular sheath.

Sixty patients (90%) developed grade 1 dermatitis and seven (10%) developed grade 2 dermatitis, with no grade 3 dermatitis. A percutaneous gastrostomy tube was placed in 14 patients (21%); they were inserted prophylactically in seven (10.5%) patients and reactively in seven patients (10.5%). The tube was removed in all nine disease-free patients, at a median 2.9 (IQR, 2.6–4.4) months from the end of treatment.

Baseline and longitudinal patient-reported outcomes are displayed in Table 4. Overall, the EORTC functional domains were stable or numerically and even statistically improved over time in both the overall and oropharynx subset, except for a transient decrease of physical function three months after completing treatment. Head and neck–specific domains such as dry mouth, sticky saliva and taste/smell (“senses”) statistically albeit modestly worsened by three months and beyond.

The composite MD Anderson Dysphagia Inventory (MDADI) score statistically improved over time, with a mean of 89.1 and 89.8 at 12 months in the whole and oropharynx cohorts, respectively. The 24-month outcomes were similar, with a mean of 92.6 for both the whole and oropharynx cohorts. Among patients with a baseline MDADI of at least 80, the 12-month scores were 92.3 and 93.4 for the entire and oropharynx groups, respectively. Comparable numbers at 24 months were 93.3 and 93.9 for the entire and oropharynx cohorts, respectively.

In this prospective phase II study of head and neck INRT, we have shown that eliminating ENI is associated with excellent oncologic control and superb PROs. The favorable results from this paradigm, especially in a patient population with heterogeneous baseline prognoses, suggest that a novel LN targeting scheme deliverable on a conventional linear accelerator may yield dramatic short- and, more importantly, long-term improvements in quality-of-life.

This study was powered with solitary elective nodal failure as the primary endpoint, because eliminating elective nodal irradiation would not be expected to increase a first primary or metastatic recurrence. Moreover, an elective recurrence synchronous with an in-field failure should not meaningfully change the disease course. Solitary nodal failures are often salvageable with a neck dissection if caught early, and neck surgery may be easier without prior neck irradiation. We only observed one patient with an out-of-field elective failure, which occurred simultaneously with a distant recurrence in a patient with very advanced locoregional disease at diagnosis. This patient received radiotherapy to all sites of recurrence and was without evidence of disease at the time of his death. Therefore, we met the primary endpoint and believe this paradigm evidenced sufficient oncologic safety to warrant further study, especially given the observed treatment plan dosimetry and quality-of-life benefits.

Indeed, the normal tissue doses achieved with this paradigm are superior to those generated using other deescalation or state-of-the-art treatments. For example, the mean superior/middle constrictor and inferior constrictor doses in INRT-AIR patients with oropharyngeal cancer were 33.4 Gy and 11.3 Gy, respectively, which are dramatically lower than the 49.7 Gy and 28.4 Gy doses obtained in the DARS study evaluating intentional dose reduction to swallowing structures (15). The oropharyngeal contralateral and ipsilateral parotid doses of 13.5 Gy and 20.9 Gy, respectively, compare favorably to 13.6 and 31.2 Gy seen with postoperative intensity-modulated proton therapy (16). Similarly, in a phase II study of total dose deescalation in favorable oropharyngeal cancer, Chera and colleagues obtained mean doses of 22 Gy to the contralateral parotid, 41 Gy to the ipsilateral parotid, 34 Gy to the larynx, and 51 Gy to the constrictors, doses all higher than those seen in this study (17).

These promising dosimetric results are a consequence of the proximity of nearly every critical OAR to the elective neck volume, including the parotid and submandibular glands, pharyngeal constrictors, larynx and esophagus. Indeed, other studies investigating elective neck dose deescalation to 30 to 40 Gy have shown favorable improvements in PRO-reported outcomes (1–3). It is therefore expected that eliminating treatment to the elective neck should lead to even further meaningful gains in quality-of-life, especially because toxicity for many OAR is a continuous function of dose. Although not measurable in a study with this time horizon, there may be additional benefits to long-term carotid and soft-tissue complications by eliminating large-field irradiation.

In fact, the superb patient-reported outcomes recorded from these patients support the translation of dosimetry gains to clinical benefit. The EORTC overall and functional scales nearly uniformly showed improvements over time. Dry mouth, sticky saliva, and taste/smell were consistently reduced after radiotherapy treatment, but because a score of 33 in these scales reflects “a little” of a given symptom, the low absolute scores in follow-up (ranging from 13.7 to 32.7 at one year) suggest the burden of these symptoms was modest. The cohort's MDADI scores compare very favorably to published data. To put a mean score of 89.8 at 12 months for the oropharyngeal cohort into perspective, the 1-year MDADI scores in ECOG 3311 (transoral surgery followed by 50 Gy for operable p16+ disease; ref. 18) and HN002 (60 Gy with cisplatin; ref. 19) were 79.1 and 85.3, respectively, despite starting with a higher baseline than this study. Indeed, among all patients in this study, starting with a baseline score of at least 80, long-term MDADI scores were more than 90. There is an inherent danger in comparing results across trials, as the patient populations were intrinsically different. Still, the outcomes of the same standardized instrument highlight the potential of the INRT paradigm to meaningfully improve the long-term patient experience.

To our knowledge, this study is the first prospective radiotherapy trial with AI-informed target contours. Although many AI models are developed, very few are used in prospective trials, and even fewer for therapeutic guidance (20–22). As the use of AI is exploding in radiation oncology (23), it is critical to prospectively interrogate its utility in clinical medicine and understand its strengths and limitations. Using this technology to classify targets may have broad value throughout many disease sites, in addition to the possibility of improving the contours themselves (24). It is difficult to estimate the marginal benefit of the treatment of the AI assigned nodes in this study. Nodes were treated due to the AI model in the clear majority of patients and in a reasonable percentage of nodal stations, and perhaps that algorithm accounts for this study's success. Because these LNs were only classified by the model after not meeting the other radiologic criteria, they would not have been intentionally irradiated. On the other hand, while the negative predictive value of our model was very high, the positive predictive value was lower, such that more nodes were treated than necessary.

Although the primary rationale of this study was toxicity reduction, the elimination of ENI may actually facilitate the integration of immunotherapy with definitive chemoradiotherapy. Immunotherapy has become the standard-of-care in the management of metastatic HNSCC (25), but prospective studies of chemoimmunotherapy have been negative (26). One compelling explanation for these negative results is that irradiation of draining LN basins eliminates the effector immune cells necessary for a successful immune response. Indeed, Darragh and colleagues elegantly showed how ENI markedly reduces immune activation in a mouse model of head and neck cancer (27). Therefore, one may hypothesize that an INRT-based treatment would mitigate the immune ablation seen with conventional head and neck IMRT, although prospective study is necessary to interrogate this hypothesis.

This study does have several limitations. First, it was a single institutional, nonrandomized trial using a proprietary AI algorithm, and confirming generalizability of this unique scheme is critical. The median follow-up after radiotherapy was beyond 2.5 years, but theoretically, regional failures may present later. Nevertheless, given that these nodal basins did not receive any intentional elective dose, it is quite unlikely additional failures will develop after two years in a meaningful number of patients, if any. Although the study cohort was composed of a nontrivial number of low-risk patients with oropharyngeal cancer, in fact the majority of patients presented with higher-risk disease. The successful outcomes thus cannot be attributed to natively favorable biology and, in fact, the heterogeneity of treated patients suggests this deescalation paradigm may be broadly applicable to HNSCC, rather than restricted to a discrete population. In addition, we chose a high dose, 66.5 Gy, for the suspicious and potentially suspicious nodes to minimize the risk of a competing “in-field” elective nodal recurrence, which would compromise the interpretation of the trial. This prophylactic dose could have led to more scattering to the surrounding nodal basins, which may have sterilized microscopic disease. However, the median elective neck V50 was 40.7%, and even the median V30 was only 57.7%, suggesting that the oncologic success of the treatment was not simply due to unintended irradiation of at-risk nodal tissue. These nodes were quite small, and the very low OAR doses and minimal acute toxicity (such as dermatitis) highlight how the dose falloff was extremely sharp. The marked reduction in dose both at the superior and inferior extent of the field, as well as medially adjacent to the dysphagia organs-at-risk, may account for the superb PRO results.

Moving forward, we are investing significant efforts to refine the accuracy of the AI model, incorporating more information than just the data held within the images of an individual LN, such as its location relative to the primary tumor and other grossly involved nodes. The long-term goal is to automate the segmentation and subsequent classification of all nodes in the neck, vastly streamlining the contouring process. Moreover, in the near future, we are planning on performing a phase II randomized trial of this INRT-AIR paradigm versus standard elective neck radiotherapy, in order to further confirm the oncologic efficacy and quality-of-life benefits of this new approach.

In conclusion, this novel radiotherapy paradigm was associated with excellent oncologic outcomes, extremely favorable dosimetric endpoints, and more importantly, impressive patient-reported quality-of-life in a heterogeneous population of patients with HNSCC. However, until more data are generated, such a target identification approach should only be performed on a clinical study with informed consent. In our institution, INRT is only delivered on a prospective randomized study of daily adaptive radiotherapy, in which both arms incorporate this INRT-AIR nodal treatment. Without significantly more prospective interrogation and experience, the routine use of this approach would be premature and risk unexpected disease recurrence. Although the elimination of ENI may be considered radical, our results suggest that such an approach may be beneficial and feasible, and we hope this study motivates more research on this concept.

D. Sher reports a patent for systems, methods, and devices for characterizing lymph nodes with a multifaceted radiomics model, pending. J. Wang reports a patent for systems, methods, and devices for characterizing lymph nodes with a multifaceted radiomics model, Serial No. 63/488,225, pending. M. Dohopolski reports a provisional patent for systems, methods, and devices for characterizing lymph nodes with a multifaceted radiomics model, Serial No. 63/488,225, pending. B.D. Sumer reports personal fees and other support from OncoNano Medicine, as well as personal fees from Intuitive Surgical and Cancer Expert Now outside the submitted work. No disclosures were reported by the other authors.

D.J. Sher: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review, and editing. D.H. Moon: Formal analysis, investigation, writing–review, and editing. D. Vo: Investigation, writing–review, and editing. J. Wang: Conceptualization, software, formal analysis, investigation, methodology, writing–review, and editing. L. Chen: Conceptualization, software, formal analysis, writing–review, and editing. M. Dohopolski: Conceptualization, software, investigation, methodology, writing–review, and editing. R. Hughes: Investigation, writing–review, and editing. B.D. Sumer: Conceptualization, investigation, writing–review, and editing. C. Ahn: Formal analysis, methodology, writing–review, and editing. V. Avkshtol: Formal analysis, investigation, methodology, writing–review, and editing.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.