A machine-learning model that first learned from unlabeled abdominal CT scans through diffusion pretraining reported much stronger organ-segmentation scores in a small benchmark, with large gains in liver boundary measures. The experiment also tested whether the approach retained performance when only part of the labeled data was used.
On liver scans, the Dice score—a measure of overlap between a predicted region and an expert annotation—rose from 0.75 ± 0.36 to 0.93 ± 0.16. HD95 fell 45%, from 11.76 to 6.50 millimeters, and average surface distance fell 66%, from 4.38 to 1.51 millimeters. The reported paired comparison had p < 5.33 × 10^-26, with Cohen’s d of 0.529.
Liver boundaries led the result
Kidney scores were also higher in the reported analysis: Dice rose from 0.90 ± 0.19 to 0.95 ± 0.10, while HD95 fell 37%. The reported p value was below 4.01 × 10^-11, and Cohen’s d was 0.271.
Across pooled multi-organ results, Dice rose from 0.78 ± 0.23 to 0.95 ± 0.07. The paper also reported 68% lower variance and a 74% improvement in boundary precision. A separate variance analysis put the reductions at 55.9% for liver, 47.3% for kidney and 68.4% for the multi-organ analysis.
At 50%, 25% and 10% labeled-data levels, the liver and kidney Dice pairs were 0.92/0.94, 0.90/0.81 and 0.89/0.71, respectively. The reported values show that liver performance held up better than kidney performance as the labeled share fell, especially at the lowest level.
Learning anatomy before segmentation
Researchers trained a denoising diffusion probabilistic model, or DDPM, without organ labels on 21 abdominal CT scans to learn structural representations. They transferred its encoder weights to downstream segmentation, then evaluated three distinct transfer strategies, including fine-tuned and frozen encoders, against random initialization.
The dataset comprised 30 abdominal CT scans from radiotherapy patients, with expert annotations for 13 anatomical structures. Training, validation and test data were split at the patient level in 70%, 15% and 15% proportions; random seeds were fixed, and the test set was excluded from model, hyperparameter and early-stopping selection.
Paired comparisons used non-parametric Wilcoxon signed-rank tests with Bonferroni correction for multiple comparisons, while Cohen’s d was used to quantify practical significance. DDPM pretraining ran for 500 epochs using Adam at a learning rate of 10^-4. Segmentation training ran for 50 epochs with batch size 8 and augmentation using horizontal flips, rotations of ±15 degrees and scaling of 0.9–1.1 times.
What the transferred features retained
One test kept the transferred encoder frozen rather than updating its weights during segmentation. The detailed liver analysis reported Dice of 0.806 ± 0.284—86.9% of fine-tuned performance, which was 0.928, and 7.3% above the random baseline, which was 0.751. The analysis notes that frozen-encoder performance was heterogeneous by organ.
To inspect what the network appeared to use, the authors applied Grad-CAM, an approximate visualization of model activation. Diffusion-pretrained encoders showed high activation in surrounding anatomical context and suppressed activation within the organ, while baseline models attended predominantly to internal organ textures. The supplied analysis cautions that these maps may reflect learned shortcuts rather than a genuine anatomical prior.
A promising benchmark, not a broad validation
The result remains bounded by the benchmark: 30 abdominal CT scans from radiotherapy patients, with expert annotations for 13 anatomical structures. The supplied analysis identifies cross-dataset testing, pathological or more complex structures and full 3D evaluation as open questions, so the reported numbers do not establish performance beyond this setup.
The work is an arXiv preprint, listed as arXiv:2608.25693v1 [cs.CV] and dated 26 Aug 2026.
Paper data and sources
Original title: Unsupervised Anatomical Feature Learning via Diffusion Models: Enhanced Medical Image Segmentation with Denoising Diffusion Probabilistic Models
Authors: Akshat G, Divyansh Gupta, Shaleen Bhatnagar et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text