2023/10/26 | Research |

Saliency feature learning for MRI

Introducing saliency feature learning to multi-sequence MRI could improve the segmentation quality of deep learning systems. This is the finding of a recent study of the Medical Image Analysis lab in collaboration with the Support Center for Advanced Neuroimaging (SCAN) and the Department of Radiation Oncology of the Inselspital, Bern University Hospital.

Overview of the proposed approach “SaRF”: During model training, for each training image, class-specific saliency maps are calculated (e.g., using Deep Taylor Decomposition during a backward process) for all potential sequence types (in this example, six) and aggregated to yield a class-distinctiveness loss. Additionally, for each class label, a feature vector from the second-to-last layer (forward pass) is extracted and compared to the corresponding class-specific saliency vector, to form a saliency-feature loss. These two terms are combined with the standard cross-entropy loss term to form the final loss function used for model training. (https://doi.org/10.1016/j.cmpb.2023.107867)

On a cohort of 2100 patient cases comprising six different MR sequences per case, the team tested a novel method that uses saliency information to guide the learning of deep learning features for sequence classification. The method showed an improvement in mean accuracy by 4.4% for segmentation. Based on feedback from an expert neuroradiologist, the proposed approach furthermore improved the interpretability of trained models as well as their calibration with reduced expected calibration error (by 30.8%).


Link to the study

Medical Image Analysis