Discriminative Consistent Domain Generation for Semi-supervised Learning

David Firmin; Guang Yang; Heye Zhang; Jennifer Keegan; Jun Chen; Raad Mohiaddin; Shu Zhao; Tom Wong; Yanping Zhang

arxiv: 1907.10267 · v1 · pith:MLLF3V2Ynew · submitted 2019-07-24 · 💻 cs.LG · cs.CV· eess.IV· stat.ML

Discriminative Consistent Domain Generation for Semi-supervised Learning

Jun Chen , Heye Zhang , Yanping Zhang , Shu Zhao , Raad Mohiaddin , Tom Wong , David Firmin , Guang Yang

show 1 more author

Jennifer Keegan

This is my paper

Pith reviewed 2026-05-24 16:49 UTC · model grok-4.3

classification 💻 cs.LG cs.CVeess.IVstat.ML

keywords semi-supervised learningdomain adaptationmedical image segmentationcardiac MRIleft atrium segmentationconsistent domain generationdouble-sided adaptation

0 comments

The pith

Double-sided domain adaptation fuses labeled and unlabeled feature spaces to enable semi-supervised cardiac MRI segmentation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a method called discriminative consistent domain generation that uses double-sided domain adaptation to merge the feature distributions of labeled and unlabeled medical images. Indirect learning is applied during adaptation to retain the information needed for a downstream segmentation task. Once the consistent domain exists, the same segmentation model can be trained on both labeled and unlabeled scans through a consistent image generation step. This is demonstrated on late gadolinium enhancement cardiac MRI from two centers for left atrium and pulmonary vein segmentation. The approach is intended to reduce the need for exhaustive manual labeling while handling distribution shifts between data sources.

Core claim

A discriminative consistent domain is produced by double-sided domain adaptation followed by indirect learning; this domain then supports consistent image generation so that a segmentation model for left atrium anatomy and proximal pulmonary veins can be learned jointly from labeled and unlabeled LGE-CMRI scans acquired at one or multiple centers.

What carries the argument

Double-sided domain adaptation that fuses feature spaces of labeled and unlabeled data, combined with indirect learning to preserve task discriminativeness and consistent image generation to train the segmentation model.

If this is right

Labeled and unlabeled images from a single center or from multiple centers can be combined for training without explicit distribution matching.
The segmentation model for left atrium and pulmonary veins can be improved by leveraging existing unlabeled scans rather than acquiring new labels.
The same generated domain supports both the adaptation step and the final task learning via consistent image generation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same double-sided adaptation plus indirect learning pattern could be tested on other segmentation targets such as ventricle or vessel boundaries in different MRI contrasts.
If the consistent domain remains task-discriminative, the method might reduce the labeling burden in multi-center studies where scanner differences are common.
The framework might be extended to regression or detection tasks in medical imaging by replacing the segmentation head while keeping the domain generation step.

Load-bearing premise

Indirect learning during double-sided domain adaptation preserves enough discriminativeness in the generated consistent domain for the segmentation task to benefit from the unlabeled data.

What would settle it

An experiment in which adding the unlabeled scans through the generated consistent domain produces no gain, or a loss, in left atrium segmentation accuracy on held-out test cases compared with training on labeled data alone.

read the original abstract

Deep learning based task systems normally rely on a large amount of manually labeled training data, which is expensive to obtain and subject to operator variations. Moreover, it does not always hold that the manually labeled data and the unlabeled data are sitting in the same distribution. In this paper, we alleviate these problems by proposing a discriminative consistent domain generation (DCDG) approach to achieve a semi-supervised learning. The discriminative consistent domain is achieved by a double-sided domain adaptation. The double-sided domain adaptation aims to make a fusion of the feature spaces of labeled data and unlabeled data. In this way, we can fit the differences of various distributions between labeled data and unlabeled data. In order to keep the discriminativeness of generated consistent domain for the task learning, we apply an indirect learning for the double-sided domain adaptation. Based on the generated discriminative consistent domain, we can use the unlabeled data to learn the task model along with the labeled data via a consistent image generation. We demonstrate the performance of our proposed DCDG on the late gadolinium enhancement cardiac MRI (LGE-CMRI) images acquired from patients with atrial fibrillation in two clinical centers for the segmentation of the left atrium anatomy (LA) and proximal pulmonary veins (PVs). The experiments show that our semi-supervised approach achieves compelling segmentation results, which can prove the robustness of DCDG for the semi-supervised learning using the unlabeled data along with labeled data acquired from a single center or multicenter studies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a routine application of double-sided domain adaptation to semi-supervised LGE-CMRI segmentation, but the abstract gives no equations, numbers, or ablations so the claim cannot be checked.

read the letter

The main point is that the paper takes an existing pattern—double-sided domain adaptation plus indirect learning to create a consistent domain—and applies it to using unlabeled scans for left atrium and pulmonary vein segmentation on LGE-CMRI from two centers. It targets a real bottleneck in medical imaging where labeled data is scarce and distributions differ between centers. That part is straightforward and addresses a practical need. The experiments are run on actual patient data with atrial fibrillation, which is the kind of setting where semi-supervised methods could help. Beyond that, not much stands out as new; domain adaptation for consistent generation in semi-supervised tasks was already in the literature by 2019, and this does not add a new theoretical piece. The soft spots are clear from the abstract alone. There are no loss functions shown, no quantitative results, no ablation studies, and no comparison to standard semi-supervised baselines. The statement that the approach achieves 'compelling segmentation results' is not backed by any numbers or error analysis, so it is impossible to judge whether the generated domain stays discriminative for the downstream task or whether the indirect learning step actually works as claimed. The central assumption that fusing feature spaces this way preserves task-relevant information is stated but not tested visibly. This paper is mainly for people already working on semi-supervised segmentation in cardiac MRI who might want to see one more application on multicenter LGE data. A reader looking for new methods or rigorous validation will find little. It does not look ready for serious peer review without the full paper supplying the missing technical details and quantitative evidence; on the abstract it would be a desk reject.

Referee Report

2 major / 1 minor

Summary. The paper proposes Discriminative Consistent Domain Generation (DCDG) for semi-supervised segmentation of the left atrium and proximal pulmonary veins in LGE-CMRI images. Labeled and unlabeled data (from one or two centers) are fused via double-sided domain adaptation to create a consistent feature space; indirect learning is used to preserve task discriminativeness, after which unlabeled data contribute to model training through consistent image generation. Experiments claim compelling segmentation performance demonstrating robustness for single-center and multi-center settings.

Significance. If the central mechanism (double-sided adaptation plus indirect learning) produces a domain that is both consistent across distributions and discriminative for segmentation, the approach could meaningfully reduce reliance on fully labeled multi-center cardiac MRI data. The setting is practically relevant given label cost and inter-center domain shift.

major comments (2)

[Abstract] Abstract (method paragraph): the claim that indirect learning 'keep[s] the discriminativeness of generated consistent domain for the task learning' is stated without any loss formulation, objective, or architectural detail; without these it is impossible to assess whether the assumption holds or whether the generated domain actually supports the downstream segmentation claim.
[Abstract] Abstract (experiments paragraph): the statement that the approach 'achieves compelling segmentation results' is unsupported by any quantitative metrics, baseline comparisons, ablation studies, or error analysis in the provided text, leaving the central empirical claim unevaluable.

minor comments (1)

[Abstract] Abstract: phrasing 'which can prove the robustness of DCDG' is imprecise; a single set of experiments cannot prove robustness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the review and the opportunity to clarify points in the abstract. We address each major comment below, noting that the abstract is a concise summary while the full manuscript provides the requested technical details.

read point-by-point responses

Referee: [Abstract] Abstract (method paragraph): the claim that indirect learning 'keep[s] the discriminativeness of generated consistent domain for the task learning' is stated without any loss formulation, objective, or architectural detail; without these it is impossible to assess whether the assumption holds or whether the generated domain actually supports the downstream segmentation claim.

Authors: The abstract summarizes the high-level idea; the full manuscript (Section 3) provides the loss formulation for indirect learning, including the task-specific objective that enforces discriminativeness during double-sided domain adaptation, along with the network architecture and training details. This objective combines a segmentation loss on labeled data with consistency constraints to ensure the generated domain remains useful for the downstream task. We can expand the abstract's method paragraph with a brief reference to the indirect learning objective if the editor permits additional length. revision: partial
Referee: [Abstract] Abstract (experiments paragraph): the statement that the approach 'achieves compelling segmentation results' is unsupported by any quantitative metrics, baseline comparisons, ablation studies, or error analysis in the provided text, leaving the central empirical claim unevaluable.

Authors: The abstract is intentionally high-level. The full manuscript (Section 4) reports quantitative metrics (Dice, Jaccard, ASD), comparisons against supervised and semi-supervised baselines, ablation studies on the double-sided adaptation and indirect learning components, and error analysis across single-center and multi-center settings. We agree that including one or two key quantitative results (e.g., Dice improvement) in the abstract would make the claim more self-contained and will revise the experiments paragraph accordingly. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The abstract describes a semi-supervised segmentation method via double-sided domain adaptation and indirect learning to produce a discriminative consistent domain, with performance demonstrated on LGE-CMRI data from one or multiple centers. No equations, loss formulations, parameter-fitting steps, or self-citations are visible that would reduce any claimed prediction or result to its inputs by construction. The central claims rest on empirical segmentation outcomes rather than a mathematical derivation chain, rendering the approach self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven premise that double-sided adaptation plus indirect learning produces a domain that remains discriminative for segmentation; no free parameters or invented entities are named in the abstract.

axioms (1)

domain assumption Double-sided domain adaptation fuses feature spaces of labeled and unlabeled data while an indirect learning step preserves task discriminativeness
Invoked in the abstract description of how the consistent domain is achieved.

pith-pipeline@v0.9.0 · 5819 in / 1231 out tokens · 20699 ms · 2026-05-24T16:49:03.805975+00:00 · methodology

Discriminative Consistent Domain Generation for Semi-supervised Learning

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)