-Because the signal intensities variety between patients/slides using thrsheolds to classify cells/vell

types is going to introduce problems-> maybe try to create an app that decides cell seg/type on training
info

-the training regions in the classification tab really do have an impact on what is called tumor and not

-After you have trained your classifier once you can keep training it more and more to make it more
accurate

Tissue Segmentation notes

-right now there seems to be big problems with he tissue segmentation

-one promising new feature seems to be dividing SOX10 by DAPI

-then maybe do some eroding

-change class by area

-dilate

-fill holes

-Tissue Segmentationv1 seems okay, lets move forward ot test out Yvonnes results

-Actually the meeting411 app may do a better job than siggue segmentation

- features are SOX10 and SOX10 median chose 31,31 for median filter*****

- the filter size feature for the SOX10 median seems to be something super important to consider- looks
like it will get rid of a lot of problems

Post processing

Fill holes on tumor

Change by area- replace not utmor with tumor min 0 max 1000- maybe does nothing- really looks like it
does nothing

Dilate 30 on tumor- 0 with 0-was anything excluded

Erode 30 on tumor- 0 with 1

Erode 30 on not tumor – 1 with 0

Dilate 30 on not tumor - 1 with 1

OOutline as ROI class 0-> tuor-**where are the ROIs coming from?
-Label class 0 to ROI Tumor

Outline ROI class 1 Stroma

-Label class 1 to ROI stroma

Changle Lab class 0

-label class 0 replace with clear

Change class 1

-label class 1 replace with clear

Tissue Segmentationv2 seems a loooot better than v1 but its still having problems with cu-00-02D 1

Investigate the normalization processes- I think this is going to be necessary later

Work with tom to figure out how to convert IM3 files ot Tiff files

There are some value judgements that need ot be made about tissue seg type which im probably going
to have talk

YOU NEED TO TRAIN ON MULTIPLE IMAGES- THIS MAY FIX A LOT OF THE PROBLEMS!!!

-TRAINING ON LOTS OF IMAGES IN THE SLIDE TRAY SEEMS OT GIVE BETTER RESULTS

CELL SEGMENTATION

Using the cellfinder app with all of the post processing steps besides the last one gives good results

Structure of cellfinder

Classification

Method

-threshold

-label 1

F1-DAPI-> median->negate->poly blobs->max, .0001

F2 dapi-> median, .3
-label2

F3 dapi->negate->poly blobs->negate->max->multiply, .3

Features

-f1

-f2

-f3

-f4

Post processing

Change by area label 1 max 3 replace with clear

Separate objects label 1 separator label 3 object dimeter 7 surroundings width 2 object heath map F3

Change surrounded label 2, surrounded by label 1 replace with label 1 min .1

Change surrounded label clear, surrounded by label 1, replace with label 1 min .1

Change label 3 replace with clear

Change label 2 replace with clear

Tissue Seg test works very well- basing lots of things on the Cell finder algorithm may give good results.

-when stains are bad the thresholding algorithms work very badly

Plan

Get CD3 and SOX10 densities for the first patient and compare-> see how bad tissue seg is

May need to have meeting with tom/regan to work on this

I need to see if training on lots of images can fix the signal disparity problem

QUESTIONS:

-there is a discrepancy in my stroma and tumor counts for CD3 that doesn’t make sense

-Answer: I think these are just cells that are overlapping my tumor/stroma boundaries

-how can we access already created apps

-how do I figure out what the types are in phenomap

-when I export the phenomap data are the numbers im seeing mean signal intensities? This doesn’t
seem to be the case

COMMENTS:

-my CD3 numbers are a lot lower than what was found in the actual data

-maybe the way these were calculated were different-ex. Only divide by the number in the
tissue or stroma

-maybe start taking a look at R to see how I’m going to process this data

-K-means clustering might be better than decision forest or baysian for tissue segmentation

-need ot mess around with change intensity function

- need ot see why exported data isn’t aligning with whats expected