Deep learning for prediction of radiation-induced oral mucositis: Need for longitudinal studies

Abstract

We read with great interest the original study, “Artificial intelligence-based prediction of oral mucositis in patients with head-and-neck cancer: A prospective observational study utilizing a thermographic approach” by Thukral et al., recently published in the Cancer Research, Statistics, and Treatment journal.[1] In this cross-sectional study, the authors described a convolutional neural network-based deep learning algorithm for classifying thermographic images of patients with head-and-neck cancer undergoing radiotherapy according to the absence or presence of early oral mucositis changes. The authors demonstrated a high accuracy (82.05%) of the proposed model for the classification of the testing dataset. We have a few comments to make regarding this study. Radiation-induced oral mucositis is a frequent complication of radiotherapy in patients with head-and-neck cancers, which can vary greatly in severity from mild erythema and pain to extremely debilitating oral ulcers precluding any per-oral alimentation.[2] The management is mainly symptomatic and may necessitate invasive means of alimentation along with interruption of radiation therapy.[3] Although the exact pathogenesis of radiation-induced oral mucositis is still poorly understood, good oral health, adequate nutritional status, and advanced modulated radiotherapy regimens have been shown to have a prophylactic effect.[4] In particular, alimentation via percutaneous endoscopic gastrostomy (PEG) early in the course of radiotherapy has been shown to prevent higher grades of radiation-induced oral mucositis and consequent interruption of therapy.[4] In this regard, the true predictive application of artificial intelligence lies in identifying those patients with head-and-neck cancer who are more likely to develop higher grades of radiation-induced oral mucositis with continued radiotherapy treatment and thus are candidates for more aggressive prophylactic measures like PEG or de-intensification of therapy. Although the study by Thukral et al.[1] showed an excellent diagnostic performance of the deep learning algorithm for the detection of early oral mucositis changes on thermography images, there were some important drawbacks. Being a cross-sectional study, the subsequent development of higher grades of radiation-induced oral mucositis with higher cumulative radiation doses could not be studied. The authors did not consider the duration, regimen, planning, and dose of radiotherapy following which the patients were evaluated. The relatively small sample size made the deep learning algorithm prone to overfitting. It is important to avoid using different images from the same patient in both the training and testing datasets—the study methodology did not mention this. The validation of the deep learning algorithm should have been done by testing on patient data collected in the natural course of their disease rather than collated enriched data. Finally, we would like to recommend that the researchers use the Medical Image Computing and Computer Assisted Intervention (MICCAI) Reproducibility checklist and Checklist for Artificial Intelligence in Medical Imaging (CLAIM) checklist developed by the Radiological Society of North America (RSNA) which provide a guide to authors and reviewers for transparent and reproducible research on artificial intelligence in medical imaging.[5-7] As many epidemiological and therapeutic factors are associated with radiation-induced oral mucositis including patient factors, concomitant chemotherapy, type of chemotherapy agent, radiotherapy regimen, and the cumulative radiation dose—there is a need for carefully collated longitudinal large datasets with patient follow-ups incorporating all these factors to develop a reliable predictive model for the risk of radiation-induced oral mucositis in patients with head-and-neck cancer. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest.