Constrative Learning for Kidney Transplant Analysis using MRI data and Deep Convolutional Networks
Leo Milecki, Vicky Kalogeiton, Sylvain Bodard, Dany Anglicheau, Jean-Michel Correas, Marc-Olivier Timsit, Maria Vakalopoulou
In this work, we propose contrastive learning schemes based on a 3D Convolutional Neural Network (CNN) to generate meaningful representations for kidney transplants associated with different relevant clinical information. To deal with the problem of a limited amount of data, we investigate various two-stream schemes pre-trained in a contrastive manner, where we use the cosine embedding loss to learn to discriminate pairs of inputs. Our universal 3D CNN models identify low dimensional manifolds for representing Dynamic Contrast-Enhanced Magnetic Resonance Imaging series from four different follow-up exams after the transplant surgery. Feature visualization analysis highlights the relevance of our proposed contrastive pre-trainings and therefore their significance in the study of chronic dysfunction mechanisms in renal transplantation, setting the path for future research in this area. The code is available at https://github.com/leomlck/renal_transplant_imaging.
Friday 8th July
Poster Session 3.1 - onsite 15:20 - 16:20, virtual 11:00 - 12:00 (UTC+2)