Can a topology-informed prompt-conditioned model universally segment uterine structures?

One model, two modalities, two label spaces

The paper asks whether a single segmentation model can handle uterine structures across ultrasound and MRI, where the two modalities differ in image appearance, anatomical context, spatial resolution, and tissue contrast. A second obstacle is label heterogeneity: publicly available datasets are designed for different clinical objectives, so one may delineate the uterus and its internal structures while another targets lesions such as fibroids. Directly pooling such datasets can cause negative transfer, where knowledge useful for one segmentation task interferes with another.

TPUS, short for Topology-informed Prompt-conditioned Universal Segmentation, is the paper's answer. It is trained jointly on a uterine ultrasound dataset (UterUS) and a T2-weighted uterine myoma MRI dataset (UMD), and it is evaluated on held-out test sets from both. The claim is not that one label space covers both modalities; it is that one model can produce each dataset's native label space.

Shared anatomy, separate input stems

The backbone separates what is modality-specific from what is shared. Each dataset gets an independent stem: a 3x3x3 convolution, instance normalization, and LeakyReLU that maps its one-channel volume to 32 channels. At each iteration, one sampled volume is routed through its corresponding stem. With the reported batch size of one, the concatenation shown in the figure simply forwards the selected stem output rather than pairing ultrasound and MRI volumes.

Everything after the stems is shared. The modality-shared six-level residual U-shaped encoder-decoder has encoder widths of 32, 64, 128, 256, 320, and 320 channels and decoder widths of 320, 256, 128, 64, and 32 channels, with resolution-matched skip connections. Pool GNN and Swin GNN layers are adapted into topology blocks at encoder levels E2-E5 and decoder levels D0-D2; high-resolution levels E0-E1 and D3-D4 keep convolutional residual blocks. Both graph modules use max-relative graph convolution to complement the backbone's local convolutional bias.

Class prompts generate each dataset's classifier

Instead of a fixed classifier, TPUS conditions prediction on text prompts. For each class in a dataset's label set, including background, the method forms a task-aware prompt such as "uterus in a gynecologic ultrasound image" or "uterine myoma in a sagittal T2 pelvic MRI image." A frozen CLIP ViT-B/32 text encoder yields a 512-dimensional embedding.

The E5 bottleneck feature and the class embedding are projected to 256 dimensions, concatenated, and passed through a shared 1x1x1 controller that outputs 153 parameters per class. Those parameters define three class-wise dynamic convolutions with channel dimensions 8 to 8 to 8 to 1. The controller is shared across datasets and classes, but the resulting filters depend on both the image and the class. CLIP embeddings stay fixed while their projection and the controller are trainable. This head replaces the fixed D4 classifier and produces native multiclass logits, so softmax yields background-inclusive predictions in each dataset's own label space.

Topology loss derives exclusions automatically

The topology-informed loss targets anatomically inconsistent predictions. The paper defines pairwise exclusion relations among a dataset's foreground labels and, unlike earlier topology-informed losses that require manually specified relations for a fixed label space, derives those exclusions automatically from each dataset's native foreground labels. Hard predictions come from the class-wise argmax of softmax probabilities; a 3x3x3 kernel then identifies violations within each voxel's 26-connected neighborhood, producing a binary critical-voxel mask.

The topological interaction term applies cross-entropy only at those critical voxels. For a single foreground class the exclusion set is empty and the term is zero; the four UMD foreground classes yield six exclusion pairs. The total loss sums Dice, cross-entropy, and the topology term across four deep-supervision outputs with weights 2^(s-1)/15 for s in 1 to 4, and the topology weight is empirically set to 10^-6. Only the full-resolution dynamic output at D4 is used at inference.

Dice differs sharply between ultrasound and MRI

On the two held-out test sets, TPUS reaches Dice scores of 0.898 for the UterUS ultrasound dataset and 0.693 for the UMD T2-weighted myoma MRI dataset, outperforming several generic and universal segmentation baselines. The paper also reports average symmetric surface distance (ASSD), with TPUS achieving the lowest ASSD among the compared methods.

The two numbers should not be read as one performance level. UterUS provides binary uterine-region masks and excludes cases with fibroids, adenomyosis, and other acquired abnormalities; UMD labels the uterine wall, uterine cavity, myoma, and Nabothian cyst when present, in patients with surgically and pathologically confirmed fibroids. Dice and ASSD are reported as mean plus or minus standard deviation over cases per foreground class, then macro-averaged, and ASSD is omitted for empty surfaces. The MRI task is therefore both multi-class and drawn from a harder, lesion-bearing population.

Training and evaluation conditions behind the scores

Both datasets follow the MultiTalent multi-dataset paradigm and its nnU-Net-based preprocessing: 1-mm isotropic resampling and z-score normalization, plus data augmentation. Training uses 192-cubed patches, batch size 1, and 1,000 epochs of 250 iterations with Nesterov SGD, momentum 0.99, weight decay 3x10^-5, and polynomial decay from 10^-3. All compared methods share splits, preprocessing, and postprocessing.

The splits are 100/10/28 cases for UterUS and 191/21/54 for UMD, for training, validation, and held-out test respectively. Because every method sees the same splits and pipeline, the reported comparison isolates the model design rather than data preparation. The held-out test sets are also small, particularly the 28-case UterUS test set, so the per-class macro-averaged scores carry case-level variability that the reported standard deviations reflect.

Where the universal claim holds and stops

The evidence supports universality across exactly the two datasets tested: one 3D ultrasound dataset with binary uterine-region masks and one T2-weighted MRI dataset with four foreground labels. The paper's own framing is that these datasets differ in modality, label space, and clinical target, and that joint learning risks negative transfer; TPUS is designed to mitigate that risk through modality-specific stems, shared graph reasoning, prompt-conditioned heads, and topology-informed supervision.

What the packet does not establish is transfer to other modalities, other uterine datasets, other label definitions, or prospective clinical use. The UterUS dataset excludes fibroids, adenomyosis, and other acquired abnormalities, so its 0.898 Dice describes a cleaner population than the UMD cohort with confirmed fibroids. The 0.693 MRI Dice is the more informative number for readers judging multi-class lesion-bearing segmentation, while the ultrasound score reflects a binary region task.

The priority is the MRI multi-class gap

The most decision-relevant comparison in this paper is not TPUS versus the baselines but the gap between its two held-out results: 0.898 Dice on binary ultrasound uterine-region masks versus 0.693 Dice on four-class T2-weighted MRI in fibroid patients. That gap is where the universal-segmentation claim is most strained, and it is the number to watch when judging whether prompt-conditioned, topology-informed design genuinely closes the distance between heterogeneous uterine tasks.

For readers weighing this framework, the practical threshold is label-space complexity rather than modality alone. A single model can serve both datasets here, but its MRI performance on uterine wall, cavity, myoma, and Nabothian cyst is the result that determines whether joint training is worth the added machinery, and it is the setting where further validation on additional datasets would matter most.

Sources used in this answer