@johnh happy to hear intrest Generally I am physician during specialization in nuclear medicine I am doing my Phd on usage of FDG PET/CT in infections of vascular prosthesis I had created only some tiny project in pytorch yet (I mean from macine learning I am programming for couple years in java, kotlin, javascript, python and now starting julia), This project I will take first all my controll group and study group to try to learn a network some segmentation of vascular prosthesis (probably I will concentrate on aorto bi femoral as those are most common) I have about 30 in study and 20 in control group plus i will add some augmantation on the rotating on Y axis and some moderate gaussian elastic deformations, maybe I will think of sth else but most other augmentation techniques that I had read will not work here , when segmentation will be done I will cut sth like 2 cm around prosthesis and try to analyze this area for patterns - probably dividing this ring like areas to parts and analyze them via 3d convolutions (I know that ussually 3d analysis is not done becouse of computation problems but I hope that strategy i envisioned would make amount of voxels small enough) Of course there will be skip connections and as additional way to improve chance of convergence I am planning to add simplified boolean 3d tensor (from area around the prosthesis) that will just tell wether a voxel is below or above a treshold (the most important information while looking for prosthesis infection is presence of gas bubbles that have far lower density than anything else)
Why this may be relevant? lack of operation when there is infection of aortic vascular prosthesis according to some authors is 100% and risk of patient death when operation is done may be as high as 50% wrong decision here makes huge diffrence, and there are not many cases of it so there are not many physicians having any experience with it.
1 Like