- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper addresses medical image segmentation with only ~1% labeled data. It proposes a reconstruction-based framework with class-specific Vectorizers that compress each image into a few coefficients used to combine learnable feature vectors shared across samples, thereby encouraging dataset-level shape and location priors. The small labeled set mainly associates each Vectorizer with a semantic class, while reconstruction on both labeled and unlabeled images provides most of the training signal. Experiments on seven medical imaging datasets show that the method is generally competitive with or better than eight semi-supervised baselines under the 1% label setting.
Critical strengths
-
The method is well motivated. Reducing annotation requirements is important in medical imaging, and using a reconstruction signal that is computed identically on labeled and unlabeled images is a sensible alternative to pseudo-labeling and consistency-based approaches, as it does not depend on the quality of a teacher model.
-
The evaluation is broad, covering seven datasets, five imaging modalities, and eight recent baselines, and is reasonably controlled. The same labeled subset and data-loading pipeline are used across all methods; the baselines are run using the authors’ official code, and the source code is provided. Notably, the proposed method achieves the lowest Hausdorff distance on all seven datasets.
Critical weaknesses
-
The entire evaluation is one table and two sentences. No standard deviations or number of random seeds are reported; at 1% labels the variance across labeled subsets is large, so single-run numbers cannot distinguish the method from the baselines. There is no ablation study, so it is unknown whether the gains come from the Vectorizer or simply from adding a reconstruction loss. No qualitative segmentation results are shown.
-
The exact number of labeled samples per dataset is not given. Several datasets are small (Drishti-GS has about 50 training images, RIM-ONE DL about 311), so "1%" cannot be realized literally. It is not stated whether sampling is per slice/frame or per patient for CAMUS, CardiacUDA and MSD Hippocampus, which changes task difficulty substantially.
-
The Vectorizer (Figure 2) computes , which is a low-dimensional bottleneck followed by a linear layer; the "shared trainable feature vectors" are the weights of that layer. The segmentation-weighted per-class reconstruction loss (Eq. 7) is the design of MONet and Neural EM, which the authors cite. Per-class decomposition with reconstruction for semi-supervised medical segmentation was already proposed in SDNet (Chartsias et al., 2019), which is not discussed. Nothing in the method is specific to the 1% regime that motivates the paper.
-
Since L_seg is only an anchor and L_rec is minimized by assigning each pixel to the component that reconstructs it best, on the 99% unlabeled images the segmentation head is effectively trained to cluster pixels by appearance, not anatomy. Structures with similar intensity but different semantics (left vs. right ventricular blood pool, optic disc vs. cup) are likely to be confused. The authors observed this collapse themselves (Section 3.5) but only patch it with a warm-up schedule, and no masks on unlabeled images are shown.
-
The 1% target is justified by clinical annotation cost, but a percentage depends on dataset size and is not a meaningful unit of annotation effort. The paper does not explain why 1% rather than a fixed small number of cases is the right target.
-
Writing can be improved. The introduction surveys general semi-supervised learning at length without connecting it to the method. The contributions appear only as three bullet points inside the related work section. Terms such as "moderate supervision regimes" and "minimally supervised" lack clarity.
Questions for authors
Q1. (W1) The reproducibility statement says all experiments were run with seed=1, i.e., a single run on a single labeled subset, while Section 4 itself states that the choice of labeled samples has a severe impact on results. Please report results over multiple seeds and labeled subsets with mean and standard deviation.
Q2. (W1) Section 3.3 states that the Vectorizer is the main novelty, but no experiment isolates its effect. Please provide an ablation that replaces the Vectorizer with a standard per-image latent vector of the same dimension, keeping everything else fixed, together with a sweep over the number of feature vectors d. Without this, the reported gains cannot be attributed to the proposed contribution.
Q3. (W2) For each dataset, how many labeled images or volumes were used? For CAMUS, CardiacUDA and MSD Hippocampus, is the 1% sampled per slice/frame or per patient?
Q4. (W2) Were baseline hyperparameters adapted to the 1% setting and to these datasets, or were the default settings for ACDC/LA at 5-10% used?
Q5. (W3) Page 5 acknowledges that the Vectorizer is “practically a matrix multiplication”, i.e., a bottleneck followed by a linear layer, yet Section 3.3 presents it as the main novelty. What is new about this module relative to a standard low-dimensional latent in an autoencoder?
Q6. (W1, W4) Please show predicted segmentation masks against ground truth and against the strongest baseline on test images from several datasets, in particular CAMUS, HC18, and ISIC-2017, where the method has lower Dice but lower HD. Please also show masks on unlabeled training images where structures have similar intensity (e.g., left vs. right ventricle), and report how often the collapse described in Section 3.5 occurs with the warmup.
Q7. (W4) In Figure 3, the four reconstructions look nearly identical despite visibly different inputs, which suggests the decoder outputs an average shape of the dataset rather than a reconstruction of each image. Can the authors show that the reconstruction actually varies with the input, for example, by comparing with a standard autoencoder trained with the same loss? How does the model handle pathological cases where chamber walls are truly absent?