The War against Nasopharyngeal Carcinoma
Nasopharyngeal carcinoma (NPC) is a unique cancer of the head and neck region that is prevalent in Southern and South-eastern Asia. It ranks the highest incidence among all smoking- and non-smoking-related head and neck cancers in Singapore. Several features distinguish the Asian variant of NPC from the rest of the world.1 In Asia, the diagnosis of NPC is invariably linked to the Epstein-barr virus (EBV) infection; it is more commonly diagnosed in men than women at a 3:1 ratio; and large-scale epidemiological studies have previously suggested an association between Asian NPC and dietary habits, including the infamous consumption of salted preserved fish during early childhood.
It is unclear if these epidemiological trends contribute to the distinct phenotype of the endemic variant of NPC, whereby these tumours are extremely sensitive to radiotherapy and chemotherapy. It is also because of this exquisite treatment response, along with the challenge of achieving wide resection margins for surgery, which supports the rationale for radiotherapy as the mainstay of treatment for NPC.
Modern radiotherapy and its impact on nasopharyngeal carcinoma
Given the integral role of radiotherapy in the treatment of NPC, it is unsurprising that dramatic improvements in the cure rates and outcomes of NPC patients corresponded with the modernisation of radiotherapy techniques in the past few decades. In the 1980s, radiotherapy was delivered using two-dimensional (2D) methods, whereby tumour targeting was crudely defined by hand-drawn borders and modulation of radiation doses (dosimetry) was grossly limited. This often led to “over-dosing” of critical normal organs and “under-dosing” of the tumour. With the transition to three-dimensional (3D) dosimetry, apart from allowing for the accurate documentation of radiation doses to tumour and normal tissues, perhaps for the first time, the radiation oncologist is now required to delineate the different structures, thus prompting a review of the traditional concepts of defining the tumour targets.
A target contouring consensus was even more imperative with the advent of intensity-modulated radiotherapy (IMRT). In contrast to 3D conformal radiotherapy, IMRT advanced the field of precision radiotherapy by a wide margin due to the ability to shape radiation doses around concave targets and producing steep dose gradients between the tumour and normal tissue interphases. Such dosimetric characteristics introduce a natural demand on target contouring accuracy. Nonetheless, it was soon evident from observational and randomised trial comparisons that the improved dosimetry with IMRT yielded superior outcomes in NPC patients – cancer control rates increased substantially (from 70% to 85-90% even for advanced stage 3-4 disease), while severe delayed complication rates such as brain necrosis, brachial plexopathy, feeding difficulties etc. were much less frequent (<10%).2,3
The unmet clinical need and appeal of automatic target contouring
It would seem counter-intuitive that the IMRT-led improvement in dosimetric accuracy introduced another major uncertainty in the radiotherapy treatment of NPC patients. During the early period of IMRT implementation, radiation oncologists had little guidance on tumour target delineation. Unsurprisingly, this resulted in wide inter-physician heterogeneity when it came to define the boundaries of the tumour.
As elegantly demonstrated in a global study involving 20 experienced head and neck radiation oncologists, Hong and colleagues revealed significant differences between physicians in terms of dose selection and tumour target delineation, so much so that none of the physicians concurred for the same patient.4 This issue is compounded by the integration of contemporary imaging modalities into the contouring process. For example, although magnetic resonance imaging (MRI) offers superior soft tissue resolution compared to computed tomography (CT) imaging in the head neck region, signal changes that are indicative of tumour invasion on MRI can be ambiguous, thereby presenting another source of inter-physician variation. Notwithstanding the fact that layering of multiple imaging datasets to outline the tumour substantially prolongs the time taken to complete this seemingly onerous task.
Naturally, with the acknowledgement of this phenomenon, it begs the question if this human-related variation bears any downstream clinical consequences. The answer would be Yes! Based on a post-hoc analysis of a randomised clinical trial in head and neck cancer patients, the Trans-Tasman radiation oncology group (TROG) investigators reported that physician contouring errors represent a major factor contributing to deficiencies in radiotherapy plan quality.5 Crucially, these flaws affect the patient’s eventual survival probability – the investigators observed a 10 percent difference in survival at two years between patients who were treated with an optimal plan versus a compromised plan, in favour of the former. These results emphasised the importance of harmonising tumour and normal tissue delineation for radiotherapy planning.
Thus, this background nicely positions the appeal of exploiting the fervent enthusiasm surrounding “artificial intelligence” (AI) to address this medical conundrum. The primary objective of developing an AI auto-contouring tool would be to create a machine with comparable ability to perform this human task. The machine can then be enhanced in subsequent phases of upgrade to repeat this task with a high level of competency, with a lesser need for human intervention.
Limitations of existing solutions
A wish list of capabilities for such a tool would include:
- comparable accuracy to the human in target contouring;
- reduce the notable inter-human variation;
- cut the time needed to outline the targets, thereby reducing demands on manpower; and
- to be able to exploit this tool for auto-contouring of structures for real-time daily treatment adaptation.
However, the development of such a tool is beset by several challenges.
Foremost, one would need to acquire a training dataset that comprises of different clinical scenarios and datapoints that are representative of the diverse clinical presentations of the disease.
Next, integrity and quality of the dataset are crucial elements to ensure the training of a robust model. In the context of imaging-related models, this is even more challenging when we consider the multimodal datasets (MRI, CT, etc.) that are required. Applying a medical brain in the model development process is also important; this is in fact a major shortcoming of most models that only consider the data without incorporating the human thinking process in distinguishing what is “normal” versus “abnormal” during image perception. Fourth, there are no well-defined benchmarks on what is considered an acceptable outcome when it comes to validating the robustness of the model. Additionally, the compatibility of inter-institutional imaging datasets is another notable point that should not be overlooked; variations in for example image acquisition protocols, and devices, however subtle, can result in a substantial drop in performance of the AI image processing tool. These factors broadly summarise and highlight the key limitations of existing commercially available tools that automate target contouring to expedite radiotherapy planning.
The current study
We therefore embarked on a study in 2015 to develop an AI auto-contouring tool for NPC.6 The advantages of such a tool are straightforward. Apart from addressing the prevailing issues as aforementioned, delineating NPC is tedious (complex cases can take up to an hour to draw) and automating this process will dramatically reduce time and improve efficiency in the clinic.
This project was a collaboration between the National Cancer Centre Singapore (NCCS) and the Sun Yat-sen University Cancer Centre (SYSUCC), Guangzhou. The study design was as such:
- we first sought to construct an AI contouring tool that possessed a reasonable level of competency (benchmarked at 70% concordance with human experts) in delineating NPC tumours;
- we assessed the accuracy of the contours across different axial sections of the tumour for example, superiorly between the skull base and the intracranial component and inferiorly between the base of tongue and nasopharynx; and
- we then tested the performance of the tool in different clinical scenarios that were representative of real-world cases for example contouring on post-induction chemotherapy versus no prior treatment exposure datasets, advanced (bulky T3-4 tumours) versus early cases (small T1-2 tumours); this allowed us to evaluate the susceptibility of the model to treatment-induced changes.
To train our model, we utilised imaging datasets that was exclusively based on MRI sequences of 818 NPC patients, and analysed them using a 3D convolutional neural network (3D-CNN). Despite the limited sample size, we leveraged on the multimodal images from the different MRI sequences (T1-weighted, T1-weighted fat suppressed, T1-weighted with Gadolinum contrast-enhanced, and T2-weighted) to increase our layers of datapoints that were fed to the 3D-CNN. This resulted in a model that produced a greater than 70 percent concordance with the designated human experts in an independent head-to-head comparison of 203 NPC cases. Among the few limitations of our model, we observed that contouring accuracy was significantly inferior to the human experts in the anatomic regions of the brain and at the junction of the nasal and oral cavity, and also for post-induction chemotherapy cases. We suspect that these flaws may be attributed to the insufficient training datapoints for these clinical scenarios.
Perhaps, the most compelling findings of our study were derived from a prospective evaluation of our AI tool among eight radiation oncologists from high-volume academic centres, who treat at least 100 NPC cases per year. In 20 test cases, the AI tool was comparable against all eight physicians in tumour contouring accuracy, achieving a 78 percent concordance with the human experts. Incredibly, when the physicians were allowed to edit an AI-derived contour for the same 20 cases following a wash-out period of six months, all but one physician demonstrated an improvement in contouring accuracy. Even more impressive was the significant reduction of inter-physician variation in tumour delineation by 54.5 percent with AI-assistance. A similar magnitude of improvement was also observed in terms of time-savings. Taken together, we have demonstrated the feasibility and quantified the gains of implementing an AI auto-contouring tool to assist with the delineation of NPC in clinical practice.
Potential impact and future work
Following the preliminary success of the present study, the next phase of this project will involve expanding the adoption of our AI tool across multiple academic centres globally. While this may seem straightforward, several queries remain unanswered:
1) How will the model perform when applied in imaging datasets that are acquired using different protocols?
2) Are we able to embed an iterative process such that the model “self-learns” from its erroneous outputs?
3) What are the learning phases and innovation adoption lag times when a new technology is deployed, notwithstanding the intrinsic nuances of the different physicians that are likely to play a major factor in this regard.
It is important to address these queries in a stepwise, systematic and thorough manner, so as to ensure the successful integration of our AI auto-contouring tool in the clinic.
In the next decade, rising cancer incidences will continue to dominate the global healthcare burden. Given that half of all cancer patients will require radiotherapy at some point of their disease, streamlining the radiotherapy workflow is paramount to ensure the optimal usage of limited radiotherapy resources in the fight against cancer.
In parts of China, where institutions can record up to 5,000 new NPC cases per year, our results suggest that implementing AI-assistance in the radiotherapy planning workflow for NPC could translate to estimated time-savings of 1,000 man-hours (41.7 days). This value-add is particularly appealing in the context of rising demands for radiotherapy and the constant lack of radiotherapy resources in low-to-middle income countries. The accuracy and robustness of our AI contouring tool can potentially change practice and transform the radiotherapy workflow in NPC and also across all other cancer types.
- Chua ML, et al. Nasopharyngeal carcinoma. Lancet 2016; 387(10022): 1012-24.
- Peng G. et al. A prospective, randomized study comparing outcomes and toxicities of intensity-modulated radiotherapy vs. conventional two-dimensional radiotherapy for the treatment of nasopharyngeal carcinoma. Radiother Oncol 2012; 104(3): 286-93.
- Au KH, et al. Treatment outcomes of nasopharyngeal carcinoma in modern era after intensity modulated radiotherapy (IMRT) in Hong Kong: A report of 3328 patients (HKNPCSG 1301 study). Oral Oncol 2018; 77: 16-21.
- Hong TS, et al. Heterogeneity in head and neck IMRT target design and clinical practice. Radiother Oncol 2012; 103(1): 92-8.
- Peters LJ, et al. Critical impact of radiotherapy protocol compliance and quality in the treatment of advanced head and neck cancer: results from TROG 02.02. J Clin Oncol 2010; 28(18): 2996-3001.
- Lin L, et al. Deep Learning for Automated Contouring of Primary Tumor Volumes by MRI for Nasopharyngeal Carcinoma. Radiology 2019; 291(3): 677-86.
Dr Melvin L.K. CHUA, Senior Consultant and Clinician-Scientist, Divisions of Radiation Oncology and Medical Sciences; Principal Investigator, Tan Chin Tuan Laboratory of Optical Imaging, Photodynamic and Proton Beam Therapy - Precision Radiation Oncology Programme, National Cancer Centre Singapore.