Teams le 1/6/2020

meet session

• soa / cancer coursera
• classification en grades
• MRI T2W <> planes
• implementation tbd
• machine avec GPU, google colab
• accès possible LISA
• lundis midi

idée de fusion de deux modèles seg+classe cf challenge Mammogramme (ref à retrouver)

https://openreview.net/pdf?id=r1xDM5DGcV

voir aussi Yaroslav Nikulin (DREAM chammenge)

Therapixel

Dear Professor,

I hope you are doing well,

As we agreed during our last meeting, the goal if this week was for me to do some research about the state of the art of my subject, which was : Predicting the Gleason Score based on mp-MRI of prostate.

As I did many online searches, I found many papers written about this subject, and that it was already treated by many different labs around the world. Specially that it was part of an online challenge.
In fact, I found different papers focusing on different aspects of prostate cancer. Some tried to localise the cancerous tissues in the MRI, some predicted the Gleason score and others tried to predict the significance of the lesions.

So all the papers I found were doing a very specific single goal job.

A new idea came to my mind, and I would like to get your feedback on it :
Instead of focusing on one thing (like predicting the Gleason Score only), what if I build a complete software, for urologues and radiologists, where they could input the MRI scan of their patient, and the software would output the MRI with notes on it, indicating :
1 - The localised lesions
2- For each of these lesions a significance (typically it would predict if the lesion is dangerous or not, if it requires biopsy or not)
3 - For each significant lesion, the Gleason Score (typically it would predict the Gleason Score of the lesions it judges as dangerous)
4 - Maybe give a general attention map (like you mentioned previously) on the general state of the patient, and what could the next steps be …

To recap, since most of the papers focus on predicting a single insight based on MRI, I would work on a software that would merge the different predictions that could be done via MRI scans, and not only focus on one prediction, but really output many predictions. Not only that would increase my learning curve, but on a technical aspect, I would research how to use the deep learning network of an insight to help predict another one (just like we discussed last week, like the mammogram thing).

What do you think? Should I aim for this ?

Thank you for your time,
Have a great day

SASSINE Naim

12:50 AM (12 hours ago)

to DEBEIR
Dear Professor,

Here is my weekly update on my work for my thesis :

I am starting off my research by finding a way to segment MRIs and detect lesions. I found many algorithms online that do that, so I am investigating the different solutions and possibilities. I also found many links to different databases that also provide more and more training and testing data.

Here are some of the links I found very useful if you would like to check them out :

https://github.com/rcuocolo/PROSTATEx_masks  
https://promise12.grand-challenge.org 
https://www.cbioportal.org/study/summary?id=prad_tcga_pub 
https://github.com/piotrsobecki/PCa-CNNs
https://github.com/PerkLab/BreastTumorDetection
https://github.com/MIC-DKFZ/PROUNET

3 challenges
prostatex,v2,promice12

existe despapers/des modèles,
lesion/score/seg prostate

existe preproc --> litterature

PC--> demande Adrien

1 UNET pour la détection de la glande
todo:
1 UNET pour la détection des régions normal/benin/malin
rédaction