The number of data points as imaging voxels and as samples in this entire dataset, including non-PCa ones, are summarized in charts above.
During the classification process, the dataset was balanced as needed and was randomly split into a training set and a test set with ratio 7:3.
The performance measurements used are the overall prediction accuracy, the precision and the recall. All results are averaged over 100 runs to guarantee their reliability.
Classification of High-Risk and Low-Risk PCa
The results show that cancerous tissues can be distinguished from benign ones with accuracy scores over 90%. However, the task of differentiating high-risk from low-risk PCa only achieves performance of accuracies around 60~70%, which is not desirable.
This is not surprising since based on the data analysis, it is very hard to observe patterns that can help the separation from the dataset, which means that it could also be hard for learning algorithms to find patterns.