Refine ROI merging

[danielsf]

The ROI merger code from #248 works reasonably well, but could still use refinement. Specifically, in areas of very faint, diffuse flux, the code tends to produce anomalously large ROIs

Ideas to try:

• Mergers are judged based on how the correlation of pixels across ROIs compares to the correlation of pixels within ROIs. Perhaps the width of the "within ROI" distribution could be used as a measure of the quality of the ROI.
• Other information theoretic measures of the distributions of pixel properties within the ROIs (entropy, mutual information between ROIs) should also be considered
• In many cases where mergers are required, lots of small ROIs cluster around large ROIs. Maybe some combination of size and total flux/correlation/etc. can be used to determine if the small ROIs are even valid.

Tasks

• [ ] Select experiments representative of prototype segmenter's performance (i.e. some "highly over segmented" some "moderately over segmented" some "reasonably segmented")
• [ ] Run ROI merger module on these experiments
• [ ] Examine statistical properties of ROIs pre- and post-merger to determine what attributes could best be used to refine merger

Validation

• [ ] Plots comparing performance of merger module before and after this work
• [ ] Possible modifications merged into merger prototype based on the above results

This is probably an open-ended research problem, rather than a ticket with a concrete deliverable

[danielsf]

As a baseline, I will include below results from the ROI merger routine (run on seed segmentation from the Feature Vector Segmenter) from the ROI merger module as it stands. These plots were generated with the following bash script. The experiments were chosen to approximately span the range of behavior vis-a-vis over segmentation.

795901850 is highly over segmented 785569447 is moderately over segmented (and has some significant vasculature shadows) 795897800 did not exhibit significant over-segmentation and probably could have gotten by without mergin.

experiment_ids=(795897800
                785569447
                795901850)

out_dir=/allen/aibs/informatics/danielsf/frog_eggs/pr
dkdir=/allen/programs/braintv/workgroups/nc-ophys/danielk/deepinterpolation
expdir=${dkdir}/experiments

export MKL_NUM_THREADS=1
export OMP_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1

echo ${exp_dir}
for exp_id in ${experiment_ids[@]}
do
    echo ${exp_id}
    log_fname=${out_dir}/${exp_id}_log.txt
    this_dir=${expdir}/ophys_experiment_${exp_id}
    echo 'running '${exp_id}
    nice python -m ophys_etl.modules.segmentation.modules.roi_merging \
    --n_parallel_workers 32 \
    --roi_input ${this_dir}/rois/deep_denoised_filtered_hnc_Gaussian_feature_vec
tor_rois.json  \
    --video_input ${this_dir}/videos/deep_denoised.h5 \
    --corr_acceptance 2.0 \
    --log_level INFO \
    --roi_output ${out_dir}/${exp_id}_rois_2sigma.json >& ${log_fname}
done

In all plots, the upper left subplot is the max projection image from the denoised movie. The upper right subplot is the correlation metric image. The lower left subplot shows ROI boundaries before merging. The lower right subplot shows ROIs after merging (colors are meaningless and just meant to highlight that two adjacent ROIs are, in fact distinct).

In the post-merger ROI plots, uncolored ROIs represent ROIs that were deemed too large to be reasonable, flagged as invalid, and escaped out of further merging. These are the cases that lead me to believe we can do some reasonable winnowing based on the width of the autocorrelation distribution within ROIs.

[danielsf]

Experiment 795901850 (highly over segmented)

[danielsf]

Experiment 785569447 (moderately over segmented)

[danielsf]

Experiment 795897800 (probably reasonably segmented on the initial pass)