AI for Head and Neck Oncology Lab

Screening and Risk Profiling

Head and neck cancer includes tumors of the lips, oral cavity, oropharynx, nasopharynx, hypopharynx, nasal cavity, paranasal sinuses, and salivary glands. Over 1 million new cases occur annually, with prognosis often varying according to the stage of disease diagnosis. Frequently, most patients are diagnosed with advanced malignancy. Head and neck cancer screening has the potential to downstage tumor diagnosis, thereby improving patients' survival rates. However, mass population screening is not recommended for patients with head and neck cancer due to disease heterogeneity and low prevalence in the general population. As such, targeted screening, which involves identifying and screening at-risk persons, is beneficial to cancer prevention and early detection.

At the AI4HNO group, we are working on improving risk profiling and screening of head and neck malignancies. Specifically, we are focusing on using AI to model multidimensional population information to determine an individual's risk level for developing head and neck cancer. Also, we are developing AI-assisted non-invasive modalities that use fluid media like saliva and blood as adjuncts for oral cancer screening and early diagnosis.

Precancer Risk Assessment

Some head and neck cancer subtypes like lip, oral, and laryngeal cancer may present with identifiable mucosal changes long before cancer onset, known as precancerous or potentially malignant conditions that are crucial to cancer prevention and early detection. These include oral potentially malignant disorders, such as leukoplakia, erythroplakia, lichen planus, lichenoid lesions, submucous fibrosis, actinic cheilitis, and epithelial hyperplastic laryngeal lesions like vocal cord leukoplakia and dysplasia. It is known that not all precancerous lesions progress to cancer, and cancer preventive therapies are recommended following the patient's risk assessment of malignant progression.

Current methods for assessing cancer risk in mucosal precancerous lesions of the head and neck are unidimensional and not personalized. For this reason, at the AI4HNO group, we are actively constructing new AI-driven risk prediction tools that integrate diverse prognostic factors for a comprehensive assessment of cancer progression among patients with precancerous lesions of the head and neck.

Cancer Diagnosis

The definitive diagnosis of head and cancer usually involves surgical biopsy and histopathology. Histologic analysis can be challenging since an experienced pathologist is required, considering tumor heterogeneity. Marked interrater variability regarding tumor diagnosis and grading, significant workload on experts to perform time-consuming manual histologic assessment of oral and maxillofacial diseases, and the risk of human error in detecting micro-invasive tumours are known limitations of current diagnostic workflow in head and neck oncology.

To mitigate these challenges, we are researching the application of automated AI-based approaches or platforms to aid pathologists in the accurate diagnosis of oral cancer using whole-slide images. In an updated workflow, such AI-based system may potentially improve the definitive diagnosis of head cancer while reducing clinicians’ workload.

Prognosis Prediction

The prognosis of head and neck cancer varies primarily with clinical subtype and stage of diagnosis. The AJCC TNM staging, which is employed to guide patient treatment, does not individualize patient selection for multimodal or targeted therapies. Likewise, one contributory factor to the overall dismal prognosis of head and neck cancer is the lack of sound and objective decision-making platforms to stratify patients according to their risk of tumor recurrence and survival probability at diagnosis or during follow-up.

At the AI4HNO group, one key focus is on applying binary, time-to-event, and causal machine learning frameworks to clinical, pathological, and molecular data of patients with head and neck cancer to develop sound AI-based tools for tumor recurrence/survival prediction and optimal treatment selection.

Clinical AI Implementation

The accuracy and validity of AI-based clinical decision-support systems even at external validation does not compulsorily translate to clinical impact and benefit following implementation. At the AI4HNO group, we are also collaborating with surgeons and oncologists to assess the impact of generic or lab-owned AI-based systems on the clinical management of patients with head and neck cancer. We are interested in conducting parallel, cross-over, or stepped-wedge trials, quasi-experimental studies, and observational studies assessing the influence of AI models on process and patient-specific outcomes in head and neck oncology.

Epidemiology

Noma is a severe gangrenous neglected tropical disease of the mouth and face that is endemic to Sub-Saharan Africa. It is one of the WHO oral diseases of public health importance. Noma often affects children between two and six years old, though disease sequelae may be evident among individuals of different ages. As a group, we are collaborating with institutions and specialist centers in Northern Nigeria, where noma incidence and burden is high, to estimate the incidence, prevalence, and mortality of noma and define its clinical presentation in the region.

Spatial Profiling

We are using noma incidence estimation methods and robust disease mapping approaches based on Bayes estimation to define the geographical distribution of noma worldwide. We have pioneered the application of this approach by defining the small-area incidence of the disease in Northern Nigeria. We hope to obtain more areal data on the disease distribution of noma globally to model the risk of noma incidence at various sub-national levels to highlight areas for targeted individual and population health promotion activities that improve prevention, early detection, and burden of noma.