Semi-supervised Breast Lesion Detection in Ultrasound Video Based on Temporal Coherence
Pith reviewed 2026-05-24 21:07 UTC · model grok-4.3
The pith
Semi-supervised temporal coherence aggregates key-frame features to detect breast lesions in ultrasound videos at 91.3% mAP.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The method aggregates features from historical key frames chosen by an adaptive scheduling strategy and uses a new WarpNet to perform both spatial warping and feature aggregation in one step. This transfers supervision from a separate collection of labeled still images to unlabeled video sequences, yielding 91.3% mean average precision at 19 ms per frame on 1,060 ultrasound sequences compared with 86.6% mAP and 32 ms per frame for a RetinaNet detector.
What carries the argument
Adaptive key-frame scheduling for temporal coherence feature aggregation, implemented via WarpNet that replaces separate spatial warping and aggregation steps.
If this is right
- Unlabeled video sequences become usable for detection once a set of labeled still images is available.
- Inference runs at 19 ms per frame on GPU instead of 32 ms.
- Mean average precision rises from 86.6% to 91.3% on the evaluated 1,060 sequences.
- The single WarpNet module simultaneously handles warping and aggregation, removing two separate operations.
Where Pith is reading between the lines
- Annotation effort could shift from costly video labeling toward cheaper still-image labeling while retaining video detection performance.
- The same coherence-based transfer might apply to other medical video tasks where only image-level labels exist.
- Real-time clinical deployment becomes more feasible at the reported 19 ms frame time.
Load-bearing premise
Labeled still images supply supervision that transfers reliably to unlabeled video sequences through temporal coherence without domain shift or annotation mismatch.
What would settle it
Performance falling below 86.6% mAP on a fresh collection of ultrasound videos when the labeled still images come from scanners or patient populations different from the test videos.
read the original abstract
Breast lesion detection in ultrasound video is critical for computer-aided diagnosis. However, detecting lesion in video is quite challenging due to the blurred lesion boundary, high similarity to soft tissue and lack of video annotations. In this paper, we propose a semi-supervised breast lesion detection method based on temporal coherence which can detect the lesion more accurately. We aggregate features extracted from the historical key frames with adaptive key-frame scheduling strategy. Our proposed method accomplishes the unlabeled videos detection task by leveraging the supervision information from a different set of labeled images. In addition, a new WarpNet is designed to replace both the traditional spatial warping and feature aggregation operation, leading to a tremendous increase in speed. Experiments on 1,060 2D ultrasound sequences demonstrate that our proposed method achieves state-of-the-art video detection result as 91.3% in mean average precision and 19 ms per frame on GPU, compared to a RetinaNet based detection method in 86.6% and 32 ms.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a semi-supervised method for breast lesion detection in ultrasound videos that transfers supervision from a separate collection of labeled still images to unlabeled video sequences. It aggregates features from historical key frames selected by an adaptive scheduling strategy and introduces WarpNet to replace traditional spatial warping and feature aggregation for improved speed. On 1,060 2D ultrasound sequences the method reports 91.3% mean average precision at 19 ms per frame, outperforming a RetinaNet baseline (86.6% mAP, 32 ms).
Significance. If the reported gains prove robust to domain shift and data-split choices, the approach could reduce annotation burden for video-based medical detection tasks by exploiting temporal coherence. The speed gain from WarpNet would be practically relevant for real-time ultrasound analysis. The work does not supply machine-checked proofs, open code, or parameter-free derivations.
major comments (3)
- [Abstract and §4] Abstract (semi-supervised setup paragraph) and §4 Experiments: the central 91.3% mAP claim requires reliable transfer of supervision from labeled still images to video sequences, yet no patient-disjoint splits, scanner metadata, statistical comparison of lesion-size or appearance distributions, or domain-similarity diagnostics between the two data sources are reported. If domain shift exists, the 4.7-point gain over RetinaNet could be an artifact of data mismatch rather than the temporal-coherence mechanism.
- [§4.3] §4.3 Results (and associated tables): mAP figures are given as single point estimates without error bars, standard deviations across runs, or statistical significance tests; the improvement over the RetinaNet baseline therefore cannot be assessed for reliability.
- [§3.2 and §3.3] §3.2 (adaptive key-frame scheduling) and §3.3 (WarpNet): no ablation experiments isolate the contribution of the scheduling thresholds or the WarpNet architecture; without these controls the performance attribution to temporal coherence remains unverified.
minor comments (1)
- [§3] Notation for the adaptive thresholds and WarpNet input/output tensors is introduced without an explicit equation or diagram reference, making the aggregation step harder to follow.
Simulated Author's Rebuttal
We thank the referee for the constructive comments. We address each major point below and will revise the manuscript accordingly where changes are feasible.
read point-by-point responses
-
Referee: [Abstract and §4] Abstract (semi-supervised setup paragraph) and §4 Experiments: the central 91.3% mAP claim requires reliable transfer of supervision from labeled still images to video sequences, yet no patient-disjoint splits, scanner metadata, statistical comparison of lesion-size or appearance distributions, or domain-similarity diagnostics between the two data sources are reported. If domain shift exists, the 4.7-point gain over RetinaNet could be an artifact of data mismatch rather than the temporal-coherence mechanism.
Authors: Both the labeled still images and the 1,060 video sequences were collected from the same hospital under similar clinical protocols for breast ultrasound, which we believe reduces the risk of substantial domain shift. The RetinaNet baseline was trained on the identical labeled image set and evaluated on the same video sequences, making the relative 4.7-point gain attributable to the temporal-coherence components within this shared setup. However, the manuscript does not report patient-disjoint splits, scanner metadata, or explicit distribution comparisons. We will add a limitations paragraph discussing these data-source details and the absence of formal domain diagnostics. revision: partial
-
Referee: [§4.3] §4.3 Results (and associated tables): mAP figures are given as single point estimates without error bars, standard deviations across runs, or statistical significance tests; the improvement over the RetinaNet baseline therefore cannot be assessed for reliability.
Authors: We agree that single-point mAP values limit assessment of reliability. In the revised manuscript we will report mean mAP and standard deviation over at least three independent training runs with different random seeds, and we will include a paired statistical significance test (e.g., Wilcoxon signed-rank) between the proposed method and the RetinaNet baseline. revision: yes
-
Referee: [§3.2 and §3.3] §3.2 (adaptive key-frame scheduling) and §3.3 (WarpNet): no ablation experiments isolate the contribution of the scheduling thresholds or the WarpNet architecture; without these controls the performance attribution to temporal coherence remains unverified.
Authors: The current experiments present only the full pipeline. To isolate contributions we will add ablation tables in the revision that (i) disable the adaptive scheduling thresholds (replacing them with fixed-interval selection) and (ii) replace WarpNet with conventional spatial warping plus separate aggregation, reporting mAP and runtime for each variant on the same 1,060 sequences. revision: yes
Circularity Check
No circularity; empirical semi-supervised method evaluated on held-out sequences
full rationale
The paper describes a semi-supervised detection pipeline that transfers supervision from labeled still images to unlabeled videos via temporal coherence and a WarpNet module, then reports mAP on 1,060 held-out ultrasound sequences. No equations, fitted parameters, or self-citations are shown to reduce the reported 91.3% mAP or speed figures to quantities defined by construction from the training inputs themselves. The central performance claim remains an external empirical measurement rather than a self-referential renaming or prediction forced by the model's own definitions.
Axiom & Free-Parameter Ledger
free parameters (1)
- adaptive key-frame scheduling thresholds
axioms (1)
- domain assumption Temporal coherence between frames is sufficiently strong and consistent to improve detection when aggregated
invented entities (1)
-
WarpNet
no independent evidence
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.