Learned Deep Radiomics for Survival Analysis with Attention

Archive ouverte

Morvan, Ludivine | Nanni, Cristina | Michaud, Anne-Victoire | Jamet, Bastien | Bailly, Clément | Bodet-Milin, Caroline | Chauvie, Stephane | Touzeau, Cyrille | Moreau, Philippe | Zamagni, Elena | Kraeber-Bodéré, Francoise | Carlier, Thomas | Mateus, Diana

Edité par CCSD -

International audience. In the context of multiple myeloma, patient diagnosis and treatment planning involve the medical analysis of full-body Positron Emission Tomography (PET) images. There has been a growing interest in linking quantitative measurements extracted from PET images (radiomics) with statistical methods for survival analysis. Following very recent advances, we propose an end-to-end deep learning model that learns relevant features and predicts survival given the image of a lesion. We show the importance of dealing with the variable scale of the lesions, and propose to this end an attention strategy deployed both on the spatial and channels dimensions, which improves the model performance and interpretability. We show results for the progression-free survival prediction of multiple myeloma (MM) patients on a clinical dataset coming from two prospective studies. We also discuss the difficulties of adapting deep learning for survival analysis given the complexity of the task, the small lesion sizes, and PET low SNR (signal to noise ratio).

• Description
• Sujet/Public
• Outil
• Outil d'anticipation
• Mémoire et thèse
• Gestion