Fig. 5: 3D tissue reconstruction with ELD.
From: Spatial landmark detection and tissue registration with deep learning
a, Illustration of the 3D modeling process. (i) Initially, select a reference image at random from the tissue stack, choose a specific tissue and create a counterpart with introduced deformations to map onto the reference. This step involves generating a deformed version of the source tissue to simulate variations or distortions. (ii) Next, detect landmarks for both the reference image and the source tissue, including its deformed version. Finally, map the deformed source image to the reference using TPS, and then align the original source image with the reference using both TPS and a rigid transformation technique. (iii) Compute the similarity loss between the mapped noisy source and the original source image, and compare the area change between the source image mapped with TPS and rigid transformation. Repeat this process for all tissues in the stack until convergence. (iv) Finally, map all tissues to the reference. b, Demonstration of final registration using ELD, and with the manual annotations for 260 prostate samples. c, Display of aligned tissues with their anchors from the tissue stack. A total of 20 landmarks were used for the alignment. d, Performance comparison of ELD and other models based on the 260 prostate samples. All results are normalized using the value obtained when aligning with manual landmarks (corresponding to a score of 1). e, Absolute error between manual alignment and alignment using ELD across four sections in the z-stack.
