Sarcomas

The transition from cows to sarcomas was not so simple. There are several differences between cow pictures and sarcoma MRIs.

The first one is the number of pixels in each pictures. For cows, we were working on 256x256 pixels, but for MRIs, the pictures had a size of 512x512. For our computations, we were working on a cluster (Plafrim, provided by Inria), because our personnal computers could not handle the size of the data. But we had a restriction in memory and computing time on this cluster, so we had to reduce the size of these MRIs down to 256x256.

The second difference is the size of the data base. For cows, we had around 130 pictures, so our data set was equal to our data base. However, for MRIs, we had around 3000 pictures in our data base. Again, because of the restriction in memory and computing time, we had to reduce the data set to a size of 1000 pictures. When choosing these 1000 pictures, we also had to make sure that all MRIs from one patients were together in either the training set or the testing set. If not, the network would come to learn to recognise patients, not how to segment pictures.

One last difference we knew would prove to be a challenge, is the fact that in these MRIs, there aren't only sarcomas, but also bones and organs. So the network will have to learn to differenciate between all these elements. We did not have this issue with cows, because every picture only consisted of a cow and some background (a field or a barn). There is also another issue : these MRIs are a succession of cross-sections from several patients, which means that in some of them, there is no sarcoma to find.

We started training our network on sarcomas with the weights we had obtained from training on cows, so as not to start from scratch and gain time. We did not perform any data augmentation, but we did keep some drop-out and random shuffling.

We have very few results on sarcomas, but here is oen that is representative of how our network performed.

We can see in the picture on the right and side that that the network started recognizing the sarcoma, but it also included the outline of the leg. Had we had the occasion of having a longuer computing time and memory allowances, we would probably have been able to improve our results, by testing different parameters as we did with cow images.