relationship between DICE and clinical metrics

[msrepo]

Since, our clinical metrics evaluation has lot of outliers (due to difficulty in automatic evaluation of these metrics), we need to filter outliers when line fitting to obtain relationship between DICE and clinical metrics, or else use L1 Regression.

Before outlier removal

with 90% quantile outlier removal

[msrepo]

This relationship is more or less similar for other generalized metrics such as ['DSC','NSD','ASD','HD95',]

[msrepo]

The above figure is for a single architecture. Do this for all the architecture in a single figure.
Also, show the comparison between methods for relationship between dice and clinical metrics.

[msrepo]

i) femur head center seem to be consistently well located as dice score improves.
ii) vertebral body seem to be well reconstructed even for comparatively low dice(0.8 vs 0.9), any improvement in dice score is correlated with better reconstruction of spinous process.

The first part of statement ii) looks obvious since such a large part of the vertebra volume consists of vertebra body. Even when any peculiarities in the vertebral body such as osteophytes are not reconstructed, dice score is not affected much. For a interesting work on osteophyte reconstruction see Ambellan where region of low variability and high variability is obtained from Statistical Shape Model(SSM) to regularize the shape