ROI identification

[john-judge]

ROI_identification.ipynb

[john-judge]

Wen-Chi's data (06_01_01.zda)

AIC/BIC in 0-50:

AIC/BIC in 50-80:

AIC in steps of 4:

AIC in steps of 6:

AIC in steps of 10:

AIC in steps of 20: Finally seems like an elbow exists around 190.

But plotting shows that something strange and likely invalid is happening at k=190:

k=140 is possibly better. But then again it doesn't qualitatively look much more useful than k=90, for this particular dataset

[john-judge]

7/12/22 Data, 01_01_09

[85, 95]

Here are the sampled points in the percentile range [85, 95]: Resulting clusters (k=85): Example traces for these clusters (now writing to region DATs for load into PhotoZ):

[95, 100]

Here are the sampled points in the 95% percentile: Resulting clusters (k=85): Example traces for these clusters:

[85, 100]

Finally, for comparison, clustering is performed on the lumping together of the previous two percentile ranges. Resulting clusters (k=85):

Method Development

It could possibly make a lot of sense to “stratify” the pixels by SNR ranges and then perform clustering on these SNR ranges separately.

Then the groups of clusters are recombined. If two clusters are touching, the one from the lower-SNR-percentile group is discarded (because the contiguous low-SNR ROIs are probably the “halo” around or between higher-SNR ROIs).

On Choice of Percentile Ranges Based on Dataset

This dataset just happened to behave nicely with filtering to 95th percentile, maybe because with the 40 uA stim, the stim site doesn’t dominate with holding all the high-SNR pixels, and the 95th percentile is more evenly distributed, both in intensity and space.

[john-judge]

Implemented "Strata" idea (mentioned in previous post) and other filters on clusters

[john-judge]

Comparison to Kate's K-means clustering (KKM)

Kate mentions that KKM is usually followed by manually selecting about half of the clusters shown above.

[john-judge]

Comparison after KKM manual filtering