MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Knowledge Discovery From Existing Radiotherapy Patient Databases Based On Unsupervised Learning Methodology

D Tewatia1*, R Tolakanahalli2, (1) University of Wisconsin Madison, Madison, WI, (2) Miami Cancer Institute, Miami, FL

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: explore an unsupervised learning methodology for harnessing knowledge from existing radiotherapy patient databases.


Methods: in radiotherapy patient databases have increased many fold in past decade. Vast knowledge in these databases can augment recent radiotherapy treatment planning efforts of knowledge based planning. In this study, spatial and texture based parameters were computed for 9 randomly selected patient’s planning target volumes (PTVs) and surrounding critical Organs At Risk (OARs). A hierarchical clustering technique was chosen to find similarity between these patients and group them in their respective clusters. Main advantage of hierarchical clustering is that there is no need to specify number clusters upfront. The number of clusters are driven based on data available in each database. This approach is perfect for radiotherapy patient databases where every patient is unique and knowledge keeps increasing with each new case added to existing databases. As first step a distance matrix was computed based on Pearson correlation methodology where correlation between pairs is computed by subtracting it from unity. Under this approach two pairs are close to each other if their feature vectors are highly correlated. A heatmap and dendrogram were generated to visualize this correlation. A dendrogram is a tree like representation to demonstrate the inter-relationship of similar feature vectors and heat map is a color coded map to identify patterns.


Results: PTVs and OARs for all 9 patients were clustered, and Pearson correlation maps were generated. Correlation index varied from 0.73 to 0.99 and an index > 0.9 showed excellent agreement with the dendrogram and heat map based visualization plots. The plots with the cross-correlation index enhances the deduction of patient data correlation with the other remaining eight patients.


Conclusion: study demonstrates that hierarchical clustering technique is a very promising technique for automated knowledge extraction from existing radiotherapy databases.

Download ePoster [PDF]

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email