rmsandu / segmentation-eval Goto Github PK

• create ellipsoid using the radii from the manufacturer ablation devices
• use the needle trajectory to find the angle around which to orientate the ablation

the ablation ellipsoid needs to be placed wrt to the ellipsoid

• Fix the graphs with the legend and same size

• extract the centroid coordinatres of the tumor
• extract the centroid coordinates of the ablation
• bug: Setting Sequence to an array. still haven't figured out how to add the tuple (x, y, z) to an array value

• LTP extraction
• bar plots percentage of surface margin distance
• scatter correlation plots
• add boxplot
• extract all the margins

• compile all the euclidean surface distances from the ablation_metrics_per_lesion.csv (second sheet)
• import LTP into the DB (follow-up columns)

• deep tumors
• subcapsular

• boxplot
• separation per subcapsular vs. non-subcapsular

• different coloring scheme (0-1 mm yellow, light green 1-5 mm, 5-10 mm green, deep green 10-15 mm)
• change the division to : optimal > 5mm, sufficient 0 <x<5 and not covered x<5
• send condition for the margin ranges as input to the function

• All Filepaths into CSV from GT 2017-2018?
• One CSV file per patient saved into the patients folder?
• The CSV should be generated automatically from CAS ONE IR Segmentation Software
• What additional information should be saved? Pathology type? Trajectory?
• Problem: sometimes multiple trajectories have been used for the same lesion

• correlation matrix between all features
• include energy
• LTP extraction bar plots percentage of surface margin distance scatter correlation plots
• add it all into one PPT
• scatter plot for ablation axis extracted in all directions (coronal, saggital, axial)

• read the invididual csv file with extracted radiomics
• pile them all together alongside the input batch processing
• do lesion match

• extract the maximum and minimum axis of the ablation volume

• add subcapsular/ deep lesions option
• plot Tumor Volume + 10 mm vs Predicted Ablation Volume
• separate colours for subcapsular/deep lesions on the PAV vs. EAV

Options for computing the patients segmentations:

• batch processing
• single patient processing

• G11
• G13

• the mask is not placed at the correct location
• direction/ origin might be wrong
• paste before re-sample and resize?
• identify where the issue appears : in DicomWriter or in Resize_Resample

• use the CSV with the paths downloaded from RedCap
• this way we filter out which patients we want to analyze. Remember we want to focus solely on 50 lesions deemed analyzable in the beginning.
• combine the RedCap file that has patient and lesion information with path references from the metatags

• predict Ablation Volume
• predict Ablation Volume using the formula
• extract the OOB error

• how many lesions actually treated percutaneously (175)
• how many available complete datasets
• how many treated with MWA
• how many treated with RFA
• how many re-ablated
• how many had multiple needles in/parellel ablations
• how many subcapsular
• how many vicinity vessels

• add log scale for tumor volume
• compute EAV vs. tumor volume+10 ml

Options to do the matching:

• use the CSV with the paths downloaded from RedCap
• this way we filter out which patients we want to analyze. Remember we want to focus solely on 50 lesions deemed analyzable in the beginning.
• combine the RedCap file that has patient and lesion information with path references from the metatags

The filepaths were incorrectly assigned in the final dataframe.

• predict major axis ablation
• predict minor axis ablation
• 2 separate RF models
• train only on non-subcapsular
• train on all lesions
• [ ] RF per device
• [ ] add lateral error.
• play with the number of trees, etc.
• test on the manufacturer's brochure
• perform interpolation

• Import the radiii for each deviice power and time into the LIT Postgres Database (Bitbucket Repo)

• Extract power, time, type of needle from MWA database
• MWA database useless, extract from REDCAP
• compute coordinates of the simulated ablation ellipsoid
• extract origin of the ablation zone
• add the needle offset
• re-create DICOM mask of the predicted ablation
• compare the simulated ablation with the resulted ablation segmentation

• plot the linear regression line
  -[ ] add the r-square correlation
• plot the number of samples in each graph

• extract formula from simeon's paper

• add argparse arguments for the input to avoid modifying the script every single time

Doesn't take into account that the optimal coverage is at a distance of [5-10mm] . Considers perfect score 1 if all the tumor is covered and residual tumor volume is 0.

• Energy
• Power
• ratio ablation volume/energy
• scatter plot

for example，I have the result of a brain tumor segmentation.But to distinguish different areas of tumor，We have 3 labels.

In this case , what should I do to use your code to evaluate the result?

Thanks!

• just resample to the same spacing and mask size.
• no need to save the image?

Resize and Resample to correctly compute the distance and the volume metrics:

• resize
• resample
• write to disk anonymized source files
• write to disk anonymized segmentation files
• better re-naming of tuple (or other data structures)

• extract TPEs from maverric (update xml-recording extraction algorithm)
• validated needles
• import function LIT

• Anonymize source data (CT from which segmentations were derived) as well
• Anonymize XML and other files

• extract the TP trajectory from the XML logs
  (time consuming, will deal with later)

• CSV/Excel filepath with all the patient that have one lesion.

• EAV vs. PAV
• Dice Score (or any other overlap measure) vs. the Lateral Target Error
• Effective largest axis ablation vs Predicted largest axis
• Effective largest axist ablation vs. Energy [kJ]

• minor axis
• major axis
• least axis
  For both tumor and ablation

• remove speckles
• fill holes
• smooth out the edges
• connect the separated parts