IADR Abstract Archives
Deep Learning to Automate Survival Prediction for Oral Cancer
Objectives: Deep learning is a state-of-the-art tool to analyze prognostic outcomes of cancer patients. However, the use of such analytic method for outcome prediction of oral cancer remains unexplored when compared with other types of cancers like lung and liver cancer. The present study aims to use deep learning to predict one-year survival outcomes of oral squamous cell carcinoma using open-sourced histology images.
Methods: This retrospective study used head and neck histology images retrieved from Clinical Proteomic Tumor Analysis Consortium. The dataset included a total of 391 scans with 188 scans of “alive within 12 months” and 24 scans of “death within 12 months”. Annotation on whole slide images was performed and verified by a pathologist in terms of cell types, colors of hematoxylin-and-eosin stain, and image quality. After annotation, 122 scans were qualified for subsequent image tiles generation, then followed by the process of normalization and augmentation. Quantitative evaluation on the performance of the deep learning model was analyzed using the area under the curve.
Results: The curated dataset contained more than 30,000 image tiles from 122 scans of tumor, dysplasia, and non-tumor tissue. With a fair split ratio (80:20) for training and test dataset, our deep learning model is expected to achieve more than 50% accuracy.
Conclusions: In summary, using curated image data in a sufficient quantity can be used to develop deep learning model for predicting survival outcomes, but may showing suboptimal performance. Future work for enhancing the model performance will be using other machine learning methods to extract additional features.
Methods: This retrospective study used head and neck histology images retrieved from Clinical Proteomic Tumor Analysis Consortium. The dataset included a total of 391 scans with 188 scans of “alive within 12 months” and 24 scans of “death within 12 months”. Annotation on whole slide images was performed and verified by a pathologist in terms of cell types, colors of hematoxylin-and-eosin stain, and image quality. After annotation, 122 scans were qualified for subsequent image tiles generation, then followed by the process of normalization and augmentation. Quantitative evaluation on the performance of the deep learning model was analyzed using the area under the curve.
Results: The curated dataset contained more than 30,000 image tiles from 122 scans of tumor, dysplasia, and non-tumor tissue. With a fair split ratio (80:20) for training and test dataset, our deep learning model is expected to achieve more than 50% accuracy.
Conclusions: In summary, using curated image data in a sufficient quantity can be used to develop deep learning model for predicting survival outcomes, but may showing suboptimal performance. Future work for enhancing the model performance will be using other machine learning methods to extract additional features.