Fig. 2: Deep learning-based framework to identify abnormal heart morphology.
From: Deep learning-based detection of murine congenital heart defects from µCT scans
a A nnU-Net is used to segment hearts from μCT scans. The 3D renderings show the entire μCT scan on the left, and a bounding box around the segmented heart on the right (with 5 voxel padding along the three axes). This bounding box is used to define inputs for the classification module in (c). b From the stack of slices in this bounding box, we extract five distinct 3D images by sampling every fifth slice, as illustrated (each color corresponds to a distinct 3D image), leading to five times more 3D training images. c The diagnosis module is a 3D CNN that takes 3D images resized to 64 × 64 x 32 voxels as input and comprises four convolutional blocks, a 3D max-pooling layer, a global average pooling layer, a dense (fully connected, FC) layer with 512 neurons and an output layer with two neurons (encoding the predicted probabilities of hearts having CHD, p(CHD) or being normal, p(normal)). Each convolutional block comprises one 3D convolutional layer and a 3D batch normalization layer. Scale bars: 0.5 mm in (a), 1.5 mm in (b).
