(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: HPV-associated oropharyngeal squamous cell carcinoma (OPSCC) can be treated with definitive radiotherapy (RT), combined chemoradiotherapy (CRT), and/or transoral robotic surgery (TORS). TORS is a novel approach that provides a comparable overall survival (OS) rate (~80%) with the potential for less toxicity and lower cost. However, approximately 33% of TORS patients require adjuvant CRT (with the highest toxicity profile) due to extranodal extension (ENE) or positive surgical margins on the final pathology. In retrospect, the decision for these patients to undergo TORS was unfavorable. This study describes an artificial intelligence (AI) algorithm, designed to identify TORS suitable candidates with low risk of ENE or positive margins.
Methods: Electronic medical records (EMR) for 84 subjects were retrospectively reviewed. Patient data included demographics, history, vital signs, and labs. Radiomics from diagnostic CTs were also utilized. After statistical pre-processing and remapping to equivalent numerical representations, three statistically significant textures were added to the analyses. Extracted data were used as an input into a convolutional neural network (CNN), with a probabilistic value for TORS being the post-surgical pathology.
The CNN (based on Google’s TensorFlow) consisted of 40 input nodes, 3 hidden layers, and 2 output nodes. A supervised learning technique was trained to mimic the treatment decision strategy. An 80/20 data split was utilized for testing and training with 10-fold cross-validation.
Results: The CNN was successfully implemented with variable parameters, allowing flexibility and easy reconfigurability. Training time was < 1 min to reproduce patient outcomes with an accuracy of ~70% with only EMR data and ~80% with EMR and radiomics data.
Conclusion: AI techniques may help in the treatment selection of HPV-associated OPSCC patients, discriminating those who would ultimately require tri-modality therapy with the highest toxicity profile. Future directions for research include expanding the database and determining specific variables in CNN decision making.
Keywords
Machine Learning, Artificial Intelligence, Radiomics
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email
