ADPKD genotype calls move to kidney point clouds: E3PN trains 163 times faster on one GPU, yet F1 reaches only 0.226
Synopsis
Inferring autosomal dominant polycystic kidney disease (ADPKD) genotype from 3D abdominal MRI had no public benchmark, and full-abdomen voxel volumes were hard to train on a single GPU. E3PN represents each kidney as a point cloud sampled from a Gaussian-mixture density and removes orientation with principal-component coordinate canonicalization, reaching F1 0.226 on 441 patients under 10-fold cross-validation, above 3D ResNet at 0.165 and e3nn at 0.163. All three methods stay near the 0.20 chance level, so the authors position E3PN as an efficient, reproducible baseline rather than a clinically deployable classifier.
Interpretation
E3PN replaces the whole voxel grid with a kidney point cloud as input to ADPKD genotype classification, letting full-abdomen 3D MRI train on a single GPU. No public benchmark existed for five-way ADPKD genotype classification from 3D MRI, and prior ADPKD imaging work focused on segmentation, total kidney volume estimation or hand-crafted features. On 441 patients with 10-fold cross-validation, E3PN reaches F1 0.226, 3D ResNet 0.165 and e3nn 0.163, while a classifier guessing by class priors has an expected macro-F1 of 0.20.
Each image keeps only non-zero-intensity voxels in the kidney region, then draws a fixed number of points in proportion to a two-component Gaussian density fitted to the left and right kidneys. A dense voxel grid must store large zero-intensity background, and cropping to a kidney bounding box barely shrinks severely enlarged truncating-PKD1 kidneys. Training and evaluation sample 1,024 points per case; the ablation drops F1 to 0.217 without density-based sampling and to 0.196 without coordinate canonicalization.
Coordinate canonicalization maps the point set into a standard frame set by the covariance matrix eigenvectors, so the classifier need not learn orientation. The equivariant network e3nn performs tensor products at every layer to preserve E(3) equivariance, with cost growing with feature order. With the same 1,024-point input, E3PN trains in 0.2 s per epoch using 0.307 GB of GPU memory, versus 10.7 s and 0.952 GB for e3nn and 32.7 s and 31.784 GB for 3D ResNet.
Perspective
Results come from 441 ADPKD patient MRIs collected at the University of Pittsburgh, with kidney regions extracted by segmentation masks, evaluated by 10-fold cross-validation over five labels: PKD1 truncating, PKD1 non-truncating, PKD2 truncating, PKD2 non-truncating and no mutation detected. The efficiency comparison ran on a single NVIDIA RTX 6000 workstation, with point-cloud models sampling 1,024 points per case. The authors propose larger cohorts or fusion with clinical covariates as the next step.
The authors state these results are not evidence that MRI can substitute for genetic testing; they quantify how much genotype signal a purely geometric model extracts from a 441-patient cohort. The F1 gain is concentrated in separating the two PKD1 subtypes, while the two PKD2 classes and no-mutation cases are rarely predicted correctly; the two PKD2 classes together comprise only 7 cases in a test fold. The authors also note that the majority-rule sign disambiguation can fail when the point distribution is exactly symmetric about a principal axis.
