Head and neck cancer researchers demonstrate the ability of a deep learning algorithm in the post-operative setting

According to researchers from the ECOG-ACRIN Cancer Research Group (ECOG-ACRIN), artificial intelligence can augment current methods for predicting the risk of head and neck cancer spreading beyond the borders of lymph nodes by the neck. A custom deep learning algorithm using standard computed tomography (CT) images and associated data provided by patients who participated in the E3311 Phase 2 trial shows promise, especially for patients with a new cancer diagnosis of the head and neck linked to the human papillomavirus (HPV). The validated E3311 dataset has the potential to contribute to more accurate disease staging and risk prediction.

Benjamin Kann, MD (Dana-Farber Cancer Institute, Brigham and Women’s Hospital, Harvard Medical School), led the study for ECOG-ACRIN. He will present the findings at the American Society of Radiation Oncology (ASTRO) annual meeting in San Antonio, Texas.

“This type of research is critical because it can help identify patients with aggressive, high-risk disease and also select appropriate patients for therapeutic de-escalation,” Dr. Kann said.

Head and neck cancers and their standard treatments (surgery, radiotherapy or chemotherapy) cause significant morbidity. They affect the way a person looks, speaks, eats or breathes. Therefore, there is great interest in developing less intense treatment strategies for patients. For example, the completed Phase 3 E3311 trial showed that low-dose radiotherapy at 50 Gray (Gy) without chemotherapy after transoral surgery resulted in very high survival and exceptional quality of life in patients at average risk of recurrence (Ferris RL. J Clin Onc. December 2021).

Dr. Kann and colleagues developed and validated a neural network-based deep learning algorithm based on diagnostic computed tomography (CT) scans, pathology, and clinical data. The source was the cohort of participants in the E3311 trial who were assessed at high risk for recurrence by standard pathological and clinical measures.

“Head and neck cancer staging is a difficult clinical problem,” Dr. Kann said. “In particular, our current efforts to identify extranodal extension via human interpretation of pretreatment imaging have generally shown poor results.”

Factors that determine the stage of the cancer include the size of the original tumor, the number of lymph nodes involved, and extranodal extension – when malignant cells spread beyond the boundaries of the lymph nodes in the neck into the tissues surrounding. In Study E3311, patients were assessed as high risk if there was extranodal extension (ENE) ≥ 1 mm. These patients were assigned to chemotherapy and high-dose radiotherapy (66 Gy) after transoral surgery.

Dr. Kann and colleagues obtained pretreatment computed tomography (CT) scans and corresponding surgical pathology reports from the E3311 high-risk cohort, as available. Of 177 scans collected, 311 nodes were annotated: 71 (23%) with ENE and 39 (13%) with ≥ 1 mm ENE.

The tool showed high performance in predicting ENE, significantly outperforming examinations performed by expert head and neck radiologists.

“The deep learning algorithm accurately classified 85% of nodes as having an ENE, compared to 70% for radiologists,” Dr. Kann said. “As for specificity and sensitivity, the deep learning algorithm was 78% accurate compared to 62% for radiologists.”

The team plans to evaluate the dataset in future treatment trials for head and neck cancer. The algorithm will be evaluated for its potential for improvement over current disease staging and risk assessment methods.

“Our ability to develop biomarkers from standard CT images is an exciting new area of ​​clinical research and gives hope that we will be able to better tailor treatment to each patient, including deciding when best to use the surgery and in whom to reduce the scope of treatment,” said lead author Barbara A. Burtness, MD.

Provided by the ECOG-ACRIN Cancer Research Group