(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 6
Purpose: In clinical cancer prognosis and diagnosis, tumor’s differentiation status is always adopted as reference from histopathology images. In this study we trained a CNN (convolutional neural network) to generate colon cancer’s differentiation heatmaps by learning histopathology images. Outcomes of colon cancer would be predicted based on differentiation heatmaps using deep learning method.
Methods: Firstly, regions of interests were annotated and labeled by our pathologists from TCGA’s (The Cancer Genome Atlas) WSIs (whole slide histopathology images). Then small patches extracted from WSI with manual annotated labels were randomly sampled from regions of interests. Finally, these small patches were sent to model for training. The well-trained model can predict WSIs and gain heatmaps of differentiation. According to heatmap’s information and patients’ overall survival data we developed a second cox-survival CNN to predict colon cancer’s outcomes. A specific heatmap would be divided into grid regions and the risk would was the second large risk of these regions.
Results: The accuracy of predicting four categories had reached 88.0%, 92.3%, 66.8% and 88.1% and overall accuracy had reached 84.3%. The multi-class AUC were 0.983, 0.963, 0.963, 0.981. The c-statistic of predicting colon cancer's risk reached 0.692. We set median risk as watershed and divided data into two groups. Kaplan-Meier curve was drawn and related log rank test’s p value was below 0.0001.
Conclusion: We developed a CNN model for predicting differentiation status of colon cancer through whole slide histopathology images which can precisely distinguish different differentiation stages on small patches and archived to impressively high ROC. Another CNN model was trained based on heatmaps generated by above model for predicting colon cancer’s outcomes which may have high potential to facilitate colon cancer’s diagnosis and perform better individual treatment plans.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
