Bridging the Gap Between Machine Learning and Clinicians Through Interpretable AI in Head and Neck Cancer Assessment
Y Wang1, W Duggar2*, T Thomas2, P Roberts2, R Gatewood2, S Vijayakumar2, L Bian1, H Wang1, (1) Mississippi State University, ,,(2) University of Mississippi Med. Center, Jackson, MS,
Presentations
PO-GePV-I-70 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: The objective of this research is to perform Extracapsular Extension (ECE) identification automatically with an interpretable machine learning technique. We investigate AI techniques guided by routine computed tomography (CT) scans to assist with ECE identification prior to surgery which is currently not possible even though it is highly important to accurate staging. Though high performing algorithms have been created previously, they have been confusing and required lymph node delineation. This model will not require lymph node delineation and will generate “heatmaps” for clinicians to demonstrate areas highly relevant to its ECE prediction.
Methods: In this research, we propose an interpretable machine learning model to detect ECE without requiring annotated lymph node information. The studied dataset includes 130 patients’ CT scans including both ECE positive and negative patients. To check the interpretability of the model, possibility map is utilized to highlight the possible ECE region. The algorithm can provide visual explanations for deep neural networks by indicating the critical regions in the image for prediction and further generate “heatmaps” which can highlight the regions that are highly related to extension. The generated “heatmaps” are then compared with the pathology to demonstrate the connection between the algorithm and potential clinical decision.
Results: The results have demonstrated that our model can identify ECE positive and negative patients. In evaluation, the proposed method has achieved test accuracy and AUC of 71.2% and 70.5%, respectively. The generated heatmaps can potentially cover the areas where extensions exist identified by radiation oncologist.
Conclusion: This research demonstrates that our proposed machine learning model can be used for ECE identification. By introducing the probability heatmap, the algorithm can identify where ECE has possibly exists. In addition, the heatmaps can be used as a tool for clinicians in the identification of ECE and guide clinical decisions.
Keywords
Image Analysis, Image Visualization
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning
Contact Email