Kidney cortex segmentation in 2D CT with U-Nets ensemble aggregation.

PURPOSE: This work presents our contribution to one of the data challenges organized by the French Radiology Society during the Journées Francophones de Radiologie. This challenge consisted in segmenting the kidney cortex from coronal computed tomography (CT) images, cropped around the cortex. MATERIALS AND METHODS: We chose to train an ensemble of fully-convolutional networks and to aggregate their prediction at test time to perform the segmentation. An image database was made available in 3 batches. A first training batch of 250 images with segmentation masks was provided by the challenge organizers one month before the conference. An additional training batch of 247 pairs was shared when the conference began. Participants were ranked using a Dice score. RESULTS: The segmentation results of our algorithm match the renal cortex with a good precision. Our strategy yielded a Dice score of 0.867, ranking us first in the data challenge. CONCLUSION: The proposed solution provides robust and accurate automatic segmentations of the renal cortex in CT images although the precision of the provided reference segmentations seemed to set a low upper bound on the numerical performance. However, this process should be applied in 3D to quantify the renal cortex volume, which would require a marked labelling effort to train the networks.

Automatic knee meniscus tear detection and orientation classification with Mask-RCNN.

PURPOSE: This work presents our contribution to a data challenge organized by the French Radiology Society during the Journées Francophones de Radiologie in October 2018. This challenge consisted in classifying MR images of the knee with respect to the presence of tears in the knee menisci, on meniscal tear location, and meniscal tear orientation. MATERIALS AND METHODS: We trained a mask region-based convolutional neural network (R-CNN) to explicitly localize normal and torn menisci, made it more robust with ensemble aggregation, and cascaded it into a shallow ConvNet to classify the orientation of the tear. RESULTS: Our approach predicted accurately tears in the database provided for the challenge. This strategy yielded a weighted AUC score of 0.906 for all three tasks, ranking first in this challenge. CONCLUSION: The extension of the database or the use of 3D data could contribute to further improve the performances especially for non-typical cases of extensively damaged menisci or multiple tears.