Medical image segmentation means drawing an exact outline around an organ or lesion in a scan. It is held back less by model design than by labels. A radiologist needs minutes to hours to trace one volume, so most hospitals have many scans and few outlines. The usual remedy adds machinery to large convolutional networks: consistency losses on unlabeled scans, uncertainty schedules, iterative relabeling. This proposal tests a different remedy that comes from the quantum machine learning literature. That literature is built around one problem: how to map a very high-dimensional input into a model with tight capacity limits. Its best classical answer is the tensor network. Each pixel is lifted into a small vector, the pixel vectors are combined by an outer product into an exponentially large joint feature space, and the enormous weight tensor over that space is stored as a chain of small factors. Patch-based medical image segmentation using Quantum Tensor Networks showed that such a model (Strided TeNet) matches a U-Net on 2D nuclei segmentation with about 1% of the parameters. That paper trained only with full, dense masks.
We ask whether this extreme parameter economy turns into label economy. The claim is that a patch-wise tensor-network segmenter loses less accuracy than a U-Net when supervision shrinks to scribbles or a handful of labeled images. The claim is also that the advantage comes from the correlation structure of the factorized chain, not merely from having few parameters. The one knob that controls that structure is the bond dimension, meaning the width of the links between adjacent factors. We propose to choose it from the label budget using the chain's own singular-value spectrum. If either part of the claim fails, a parameter-matched tiny CNN will reveal it, and that control is built into the design.
Core Claim
Strided TeNet labels each $8\times 8$ patch with a function $$f_c(x)=\sum_{s_1,\dots,s_N} W_c^{s_1\cdots s_N}\,\phi(x_1)_{s_1}\cdots\phi(x_N)_{s_N},\qquad \phi(x_j)=\big[\cos(\tfrac{\pi}{2}x_j),\ \sin(\tfrac{\pi}{2}x_j)\big].$$ Here $x_j$ is the intensity of pixel $j$ among the $N=64$ pixels of a patch, $\phi$ lifts each pixel to two dimensions, and $W_c$ is the weight tensor for output class $c$. Stored densely, $W_c$ would have $2^{64}$ entries. Instead it is written as a matrix product state (MPS), a chain of $N$ small three-way tensors joined by links of size $\chi$, the bond dimension. That gives roughly $2N\chi^2$ parameters per class.
Two properties matter under scarce labels. First, the model is linear in each factor once the others are held fixed. Training can therefore sweep along the chain and fit one factor at a time, and each local fit is a small, well-conditioned problem even with few labeled pixels. Second, merging two neighbouring factors and splitting them again by singular value decomposition shows how much correlation the data actually supports at that link. Small singular values can be cut off.
We propose to cut them using a held-out set of labeled pixels. With only scribbles available, the chain should then stay narrow where the labels cannot justify width. A CNN has no comparable built-in, data-driven capacity dial.
The falsifiable prediction has three parts:
- On ACDC cardiac MRI, NCI-ISBI prostate MRI, and MO-NuSeg nuclei, the gap between scribble-trained and fully trained models should be smaller for the label-budgeted MPS than for a U-Net.
- The same should hold as the labeled fraction falls.
- A parameter-matched tiny CNN should not close that gap, and neither should an MPS with $\chi=1$. A $\chi=1$ chain treats pixels as independent and keeps the lift with no correlations.
Positioning
The quantum-encoding literature supplies the motivation and the constraints. Angle encoding needs one qubit per feature (QSpeech), which forces lossy reduction before encoding (Integration and Evaluation of Quantum Accelerators). Amplitude encoding packs $2^n$ values into $n$ qubits (A Data-Encoding Approach to Quantum Federated Learning), but SLIQ notes it needs circuits far too deep for current hardware.
Three results argue against putting a real quantum circuit into a segmenter now:
- Predictive Performance of Deep Quantum Data Re-uploading Models finds that deeper re-uploading circuits generalize worse, which is the wrong direction for small cohorts.
- ATP fits qubit budgets by pruning image regions below an intensity threshold. That risks deleting the faint, low-contrast edges segmentation depends on, so we do not prune.
- The COVID-CT hybrid of Hybrid Quantum Machine Learning Assisted Classification of COVID-19 shows the classical "dressing" layers carry most of the parameters. That makes the quantum contribution hard to identify. It also only classifies; it does not segment.
We therefore work entirely with classically simulated tensor networks. These are the part of this literature that already segments images. We make no claim about quantum hardware.
An end-to-end trainable hybrid classical-quantum classifier shows MPS reduction beating PCA. Theoretical error performance analysis for variational quantum circuit based functional regression gives representation bounds for tensor-train compression. We reuse the MPS machinery and the Strided TeNet architecture directly. Our change is the training signal (scribbles and few labels) and the capacity rule, not the architecture.
On the label side, two pipelines represent the alternatives we compete with:
- Learning to Segment Medical Images with Scribble-Supervision Alone grows scribbles into pseudo-masks with a random walk, a U-Net, conditional-random-field refinement, and dropout-based uncertainty filtering.
- DyCON trains a teacher–student pair on unlabeled volumes, using entropy to schedule how strictly predictions must agree.
Both put their effort into large networks plus label-propagation or consistency machinery. Both inherit the failure modes the file records: seed errors that compound over rounds and teachers that absorb students' mistakes. We test whether a structurally constrained model needs less of that machinery.
One prior result constrains the scope. Strided TeNet drops from 0.69 to 0.63 Dice against a 3D U-Net on BraTS, because flattening 3D patches into a 1D chain breaks continuity across slices. We therefore restrict the main study to 2D slice segmentation, where TeNet already matches U-Net. 3D is left as an explicit non-goal.
Research Program
Training objective. We build the label-budgeted MPS segmenter on the released Strided TeNet design and train it with partial cross-entropy on scribbled pixels only. No random walk and no conditional random field are used. We add one cheap unlabeled-data term: two intensity-jittered copies of each unlabeled patch should receive the same prediction. Because the model has only thousands of parameters, this costs little memory.
Capacity rule. Training alternates gradient steps with sweeps along the chain. On each sweep, adjacent factors are merged and re-split by singular value decomposition, keeping the smallest bond dimension whose truncation does not hurt held-out scribble-pixel loss. The reported bond-dimension profile becomes a diagnostic in its own right. We expect it to widen as labels are added, and if it does not, that is informative.
Supervision regimes. We run three regimes:
- Scribbles generated identically for every method by skeletonizing ground-truth masks.
- Full masks on 1%, 5%, 10%, and 25% of training images, with the rest unlabeled.
- Full supervision as the ceiling.
We report Dice, the 95th-percentile Hausdorff distance (HD95), parameter count, and peak GPU memory.
What is analyzed. The main object is the label-efficiency gap: the fully supervised score minus the scarce-label score, plotted per method across budgets. We also check two things directly. First, whether boundary pixels are where U-Net and MPS differ most, by computing Dice restricted to a narrow band around the true contour. Second, whether the hybrid TeNet variant's 2-layer CNN front end, rather than the chain, accounts for any gain. This mirrors the dressing concern from the quantum classifiers.
Baselines
Naive reference. The do-nothing reference is a standard 2D U-Net trained with partial cross-entropy on the same scribble pixels or the same labeled subset, with the same intensity augmentation. It has no pseudo-labels and no consistency term. It shows how much any method gains over simply ignoring unlabeled pixels.
Strongest scribble method. This is the scribble-only pipeline of Learning to Segment Medical Images with Scribble-Supervision Alone. It reports 88.0% Dice on ACDC against 91.2% with full supervision, and 78.1% against 81.8% on NCI-ISBI prostate. Those numbers came from that paper's own scribbles, so we rerun its pipeline on our generated scribbles and report both our reproduction and the published figure. Beating it means one of two things: higher Dice on ACDC or NCI-ISBI with our scribbles, or equal Dice at a smaller gap to full supervision, using roughly 1% of the parameters and with no relabeling rounds.
Strongest few-label method. For the labeled-fraction regime, the strongest competitor is DyCON, which reports 61.48% Dice on ISLES'22 with 5% labels, against 53.53% for the prior best (BCP). That is a 3D stroke benchmark outside our 2D scope. We therefore run DyCON's losses on a 2D U-Net on our datasets and give it the same tuning budget we give the MPS. MisMatch serves as a second consistency baseline. It reports a 24% improvement over supervised training on CARVE with 5 labeled slices. On MO-NuSeg, the fully supervised reference is Strided TeNet at 0.70 Dice with 5.1K parameters, matching U-Net.
Fairness. Every method gets the same data splits, the same scribbles, and the same augmentation. Each gets an equal number of hyperparameter trials, chosen on the same validation images, and the same wall-clock training cap on one GPU.
Ablations. Three ablations isolate the key ingredient. First, a tiny CNN matched to the MPS's parameter count tests whether "few parameters" alone explains any gain. Second, an MPS fixed at $\chi=1$ keeps the lift but removes all correlation between pixels. Third, an MPS with a fixed, hand-chosen bond dimension, in place of the budgeted rule, tests whether the capacity rule matters. The proposal's central claim holds only if the full model beats all three at low label budgets.
Why not just…?
Why not just train a tiny CNN with the same parameter count? This is the most serious threat. If low capacity is the whole story, a 5K-parameter CNN will be as label-efficient as the MPS, and the tensor structure adds nothing. We do not argue around this; it is the decisive ablation. A positive result requires the MPS to show a smaller label-efficiency gap than the matched CNN across at least two of the three datasets. A negative result would still be useful: it would tell the quantum-inspired segmentation community that its parameter savings are generic, not structural.
Why not just use DyCON or another consistency method on a full U-Net? These methods already narrow the gap with few labels, and on our data they may simply win on Dice. The file also notes their costs: dual forward passes, memory-heavy queues, and empirical schedules, plus teacher drift on low-contrast tissue. Our experiments report Dice and HD95 together with memory and parameters, so the outcome is a measured trade-off, not a claimed dominance. We also test the combination, adding DyCON's consistency loss to the MPS. If that combination beats both alone, the two ideas are complementary.
Why not just refine U-Net predictions with a dense conditional random field? The evidence in the file runs against this as a complete answer. Two papers are relevant. Weakly Supervised Lesion Co-Segmentation on Ct Scans found that the post-processing field improved edge precision but cut recall, while an integrated decoder reached 90.3% Dice on DeepLesion. The scribble paper needed uncertainty filtering on top of the field to stop errors compounding. We still include U-Net plus a dense field as an extra baseline in the scribble regime, because it is cheap to run.
Proposed Experiments
Main experiment. The main experiment produces label-efficiency curves on ACDC, NCI-ISBI 2013, and MO-NuSeg. Methods are the naive U-Net, the reproduced scribble pipeline, 2D DyCON, MisMatch, the matched tiny CNN, the three MPS ablations, and the full label-budgeted MPS. Regimes are scribbles, then 1%, 5%, 10%, and 25% labeled, then full masks. Each setting uses three seeds.
First decisive check. The first decisive check is whether, at 1–5% labels and with scribbles, the full MPS's drop from full supervision is smaller than both the U-Net's and the matched CNN's, beyond seed variance.
Second decisive check. The second is whether that advantage disappears at $\chi=1$. If it does, correlations between pixels in the lifted space are doing the work.
Third decisive check. The third is whether the learned bond-dimension profile grows with the label budget. If it does, the capacity rule is responding to data, not to noise.
Secondary analyses. The boundary-band Dice analysis tests whether gains concentrate at organ edges. Running the hybrid TeNet with its CNN front end frozen versus trained tests whether the chain or the front end carries the effect.
Outcomes. If the first two checks pass, tensor-network segmenters become a credible label-efficient option for 2D clinical tasks, at a fraction of the usual memory. That would also motivate solving the 3D continuity problem. If they fail, we will have shown cleanly that the economy of quantum-inspired segmentation does not extend to scarce labels.
References
- Patch-based Medical Image Segmentation using Matrix Product State Tensor Networks
- QSpeech: Low-Qubit Quantum Speech Application Toolkit
- Integration and Evaluation of Quantum Accelerators for Data-Driven User Functions
- A Data-Encoding Approach to Quantum Federated Learning: Experimenting with Cloud Challenges
- SLIQ: Quantum Image Similarity Networks on Noisy Quantum Computers
- Predictive Performance of Deep Quantum Data Re-uploading Models
- ATP: Adaptive Threshold Pruning for Efficient Data Encoding in Quantum Neural Networks
- Hybrid Quantum Machine Learning Assisted Classification of COVID-19 from Computed Tomography Scans
- An end-to-end trainable hybrid classical-quantum classifier
- Theoretical error performance analysis for variational quantum circuit based functional regression
- Learning to Segment Medical Images with Scribble-Supervision Alone
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation
- MisMatch: Calibrated Segmentation via Consistency on Differential Morphological Feature Perturbations With Limited Labels
- Weakly Supervised Lesion Co-Segmentation on Ct Scans