Due to use of sensitive data, this repo isn't being updated. The work is being done on private repo. Expect updates soon, once we've added censoring to photos
Applying AI to predict PWS treatment's effect
- input: photo before
- output: prediction (photo after)
PWS(Port-Wine Stains) is a birthmark in which swollen blood vessels create a reddish-purplish discoloration of the skin.
- PWS is treated with use of laser. Patients attend multiple sessions, that have ussualy 1-3 months break in-between them
- PWS treatment most often doesn't fully remove PWS, but it makes it significantly better
- PWS treatments takes from 1 to even 20 visits, ussualy around ~8.
- We measure patient's improvement (treatment's efficacy) using a metric we've introduced in previous studies called GCE. GCE takes into account 2 variables: area of PWS & colour of PWS (colour improvement is a big part of PWS treatment)
- In our recent research, we've found that PWS worsens overtime when treatment is stopped.
- More statistics to be provided
- Our input data are images of patients.
- We're focusing on patient's with PWS on head & neck
- Currentely we've cleaned up data only for before 1st visit and after last visit (representing GCE min & GCE max)
- Can possibly also clean up data for other visits (not only last and first).
- Data is generated via taking 6 photos of patient(scanning machine) from 6 different angles. Data is then transformed to a 3d object (aka we can move patient's head around)
We've got three possible ways of using the data:
- Use the 6 source photos
- Use the 3d "object" (for lack of better words at 2am)
- Use the snapshots of 3d "object" (aka rotate the head 50 times by 1 angle and take a screenshot of what we see)
Furthermore, we're also given the GCE (measurement of absolute improvement) for each of the photos
We've got data of 56 patients. There are only 2-3 of those specific PWS laser & measuring machines in Poland, but further cooperation with respective clinics to get more data might be possible
We'd like to accompolish either of these:
- Predict how patient's PWS will improve at the end of the treatment (or after 1 session - but this is much harder, especially for later sessions which tend to be less effective). Ideally we'd like to have a great prediction, but even a rough prediction would be helpful for patients (perhaps generate a range of photos - as to how patient could possibly improve)
- Automate GCE metric generation based on the photo (less exciting, but also useful)
As we're working with ANN, having the best possible input data, is the best way to ensure quality of our AI. Therefore following have to be done/tried out:
- Data augmentation
- Typical imagining Data Augmentation methods, like noise, cropping, rotation etc.
- (Infinite) many rotations of 3d images.
- Isolate patient's head & neck from their clothes
- Underline PWS with imagining methods (make it stand out more compared to rest of the body, or do the opposite - make rest of the body grayer)
To mind come 2 following AIs:
- GANs (especially CycleGAN)[paper] - (https://arxiv.org/abs/1703.10593)
- CNNs
I'll also be researching if there are any ANN's that deal well with 3d images, or if there were any similar applications in the field (predicting output of some treatment as a photo).
Test on un-seen data. Possible further tests offline in real life case scenario