brightside51 / metabreast Goto Github PK

The project aims to target the deformation of breasts during MRI screenings as a way to better understand and detect tumours and their exact locations, thus allowing for a safer removal if need be. This will involve using fusing multimodal data with generative, physics-based Diffusion Models.

Python 60.13% Shell 0.10% Jupyter Notebook 38.95% C++ 0.06% Cuda 0.76%

metabreast's People

Contributors

Watchers

metabreast's Issues

Revamped Training Script (Pytorch Lightning)

Video Structural Similarity (vSSIM) Loss Implementation

Modified Training Evaluation

The idea is to include more and more diverse evaluation metrics so as to ascertain which ones quantize the quality of the images the best and possibly change the direction of training so as to produce better-looking and more realistic images, thus solving thed issues with the previously observed baseline results.

Chosen Variables

Breast Volume (requires Segmentation) ;
Breast Density (Class Variable according to BI-Rads) ;
Mass vs Non-Mass Lesion (Round/Solid vs Infiltrative but still somewhat Solid) ;
Tumour Size (according to Staging T1, T2 or T3) ;
Tumour Location (5 Positions = 4 Quadrants + Central) ;

One idea would be to merge the tumour size and location by encoding it somehow.

Proposed Variables

ACR Classification of Breast Pattern (Duke Only)
Tumour Position / Quadrant (Duke Only but in Confusing Fashion)
Lesion Nodularity (???)
????

Found Information

Duke Breast Cancer Dataset

ACR / BI-RADS Classification of Breast Pattern (for 50 Patients and according to 3 Radiologists)
Imaging Features
- Breast Volume
- Tumour Acquisition Features
Clinical Features
- Lymphadenopathy or Suspicious Nodes
- Tumour Characteristics (Mol Subtype, Staging, etc)
- Tumour Grade and Location (Left or Right)
- Nottingham Grade
Race / Ethnicity
Very Inconsistent / Empty Tags

BCCT Dataset

Extreme Amount of Unknown Tags

Mutual Information

Patient Weight
Pregnancy Status
Patient Position