List of Papers Browse by Subject Areas Author List
Abstract
Tuberculosis (TB) remains a significant global health threat. The detection of Mycobacterium tuberculosis in sputum smears is an effective approach for TB diagnosis; however, traditional manual examination is laborious, time-consuming, and susceptible to inter-observer variation. Furthermore, the absence of annotated datasets has hindered the application of efficient deep learning-based models in this domain. To address this gap, we introduce Det-Y—a multi-center dataset and benchmark designed to facilitate the development of efficient deep learning frameworks for detecting Mycobacterium tuberculosis in sputum smears. Det-Y comprises 5,444 annotated bacilli from 1,970 sputum smear images collected across two clinical centers. In addition, we propose YOLO-Y, a lightweight model optimized for accurately detecting slender tubercle bacilli. Its core innovation is the Adaptive RoiStrip Block, which employs strip-shaped convolutional kernels to enhance spatial perception along both horizontal and vertical orientations, integrated with PANet for multi-scale feature analysis. Experiments on the Det-Y dataset demonstrate that YOLO-Y achieves superior detection accuracy while maintaining a lightweight architecture with low computational cost. The complete Det-Y dataset and corresponding code are publicly available at https://github.com/FanFan902/Det-Y.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4175_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/FanFan902/Det-Y
Link to the Dataset(s)
Det-Y dataset: https://github.com/FanFan902/Det-Y
BibTex
@InProceedings{XieZij_DetY_MICCAI2026,
author = { Xie, Zijian AND Liang, Yongjie AND Huang, Yawen AND Wu, Xian AND Wei, Bizhong AND Li, Yuexiang},
title = { { Det-Y: A Multi-center Dataset and Benchmark for Efficient Detection of Mycobacterium Tuberculosis in Sputum Smears } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces Det-Y, a multi-center dataset for detecting Mycobacterium tuberculosis bacilli in sputum smear microscopy images, and proposes YOLO-Y, a lightweight detector that augments YOLOv11 with an Adaptive RoiStrip Block built from sequential 1×k and k×1 strip convolutions to better capture the elongated morphology of bacilli. Experiments compare YOLO-Y against several YOLO variants and include an ablation showing modest but consistent gains from the proposed block; a subset of Det-Y is released, with full data/code promised upon acceptance. The authors position the contribution as enabling efficient, deployable tuberculosis screening in resource-constrained settings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- A curated, multi-center dataset for bacilli detection fills a real gap in the community, where publicly annotated TB smear detection resources are scarce.
- The emphasis on lightweight design aligns with deployment needs in low-resource clinical environments.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Improvements of the proposed model are modest or tied for best on Center-1; on Center-2, YOLO-Y is not best on mAP@0.5.
- Generalization is overstated: Center-2 is used for its own train/val/test rather than as a held-out center for cross-center evaluation (train on Center-1, test on Center-2, and vice versa). The claim of “strong generalization” to Center-2 is therefore not supported.
- Data splitting appears to be at the patch level with overlap, not at the image/slide or patient level. This is a critical threat to validity: overlapping crops from the same source image can leak visual patterns across splits, inflating performance.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
For revision, I would recommend: 1) Revise the statement about generalization or report cross-center generalization (e. g. , train only on Center-1, test only on Center-2; and the reverse), alongside the within-center results. The generalization claim across centers is not demonstrated, as Center-2 is used for within-center training/testing rather than held out. 2) Clarify the size of YOLO11 for the proposed YOLO-Y implementation, YOLO11n? 3) Specify the sizes of all the models in Table 1, such as YOLOv5n, YOLOv6-N, YOLOv10-N, YOLO11n, and YOLOv12n. 4) Add experimental results of RT-DETR-R18, RT-DETRv2-S, RT-DETRv3-R18, RT-DETRv4-S, YOLOv13-N and YOLO26n if possible (only papers of RT-DETR and RT-DETRv3 have been published). 5) The color for colorlinks, linkcolor, citecolor and urlcolor is in blue. 6) Reference list issues. ① The references should strictly follow Springer Reference Style, referring to the published papers in the MICCAI proceedings. ② For YOLOv12, the reference and citation of the published paper are enough, and those of the code repository are not needed. ③ The references should be up to date, e. g. , YOLOv9, YOLOv10, and YOLOv12 have been published, not preprints. ④ Citation and reference for PANet are missing.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work addresses an important and impactful application—automated detection of tuberculosis bacilli in smear microscopy—by contributing a multi-center dataset and an efficiency-oriented YOLO variant with a directional, strip-convolution module tailored for slender targets. The approach is sensible and the engineering is practical, with small but consistent gains in the presented ablations and a reasonable emphasis on deployability.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I consider the shortcomings of the proposed model pointed out by Reviewer 2, but it does not account for the dataset’s contribution, as the paper is titled “A multi-center dataset and benchmark”. Therefore, I maintain my rating for this paper.
Review #2
- Please describe the contribution of the paper
1.The paper’s main contribution is the introduction of Det-Y, a multi center annotated benchmark for detecting Mycobacterium tuberculosis bacilli in sputum smear images, and an experimental evaluation of a lightweight YOLOv11 based detector augmented by an Adaptive RoiStrip Block.
2.The dataset and benchmark constitute the primary contribution of the work.
3.The proposed model represents a modest, application specific architectural adaptation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The problem itself has clear practical clinical value. Automated detection of Mycobacterium tuberculosis in smear microscopy is highly needed in resource-limited settings, where annotated public datasets remain scarce.
The paper curates data from two centers with different staining protocols, which contributes to resource development in this domain.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.I doubt the novelty of the proposed approach. The Adaptive RoiStrip Block appears to be largely a combination of existing standard components. It does not seem fundamentally different from prior strip-based designs, and the use of different kernel sizes at different feature levels is more consistent with a multi-scale engineering adaptation than with a genuine conceptual innovation. In addition, the novelty of the learned scale parameter also appears limited.
2.Many of the reported performance gains are very small (0.001 to 0.004), and only single experimental results are reported, without any indication of variability across different random seeds. Therefore, it is difficult to judge whether these differences reflect genuine improvements or simply experimental fluctuation. In addition, the trade-off between improved recall and reduced precision is not discussed in sufficient detail, although this trade-off is important in practice. For example, on Center-1, YOLO-Y achieves an mAP@0.5:0.95 of 0.573, which is very close to YOLOv11 and YOLOv12.The paper emphasizes the higher recall, but the precision drops to 0.827, and similar patterns can be observed elsewhere, giving the impression that positive effects are emphasized while negative effects are downplayed.
3.The paper describes a procedure in which overlapping image patches generated from the same source image are randomly partitioned into the training and test sets.
4.All comparative experiments are limited to YOLO-based models. This is understandable if the goal is only to establish a YOLO-based benchmark. However, the presentation gives the impression that the proposed approach is intended to have broader effectiveness. The lack of comparisons with non-YOLO detectors may therefore lead to overstatement of the claims. In addition, although the model is repeatedly described as lightweight, only GPU memory usage is reported. This measure depends on factors such as batch size and input resolution and is therefore not sufficiently informative. More meaningful efficiency metrics, such as FLOPs, parameter count, and inference latency, are needed to support these claims.
5.Since the dataset is positioned as a major contribution, the annotation protocol, basic distribution statistics, inter-annotator agreement, and external validity should be described in more detail. Without this information, the dataset contribution is less convincing than claimed.
6.In addition, there are several writing and typography issues, including the misspelling of “evaluation metrics,” an incorrect description around Eq. (3) (describing the horizontal receptive field as vertical), inconsistent bibliography formatting, and basic typos, such as describing the change from 0.924 to 0.940 as an improvement of 0.16 when it is actually 0.016.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper suffers from methodological flaws, insufficient dataset information, and doubts regarding data partitioning and annotation consistency, which weaken the reliability of the reported performance improvements. In addition, the lack of repeated experiments, variability reporting, or statistical significance analysis makes it difficult to assess whether the reported gains are substantial. The ablation results suggest limited and somewhat inconsistent improvements; however, the impact of the proposed module is emphasized rather strongly in the paper. Although the model is repeatedly described as lightweight and efficient, the paper reports only GPU memory usage. Since GPU memory usage depends on factors such as input configuration, these claims are not convincing without also reporting more informative efficiency measures such as the number of parameters and FLOPs.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors honestly acknowledge that the Adaptive RoiStrip Block is primarily an engineering-oriented adaptation. However, the additional 5-run statistics show that the improvement on Center-1 mAP@0.5:0.95 is only 0.573±0.002 vs 0.571±0.003, which remains marginal, with overlapping confidence intervals and unclear practical relevance at this magnitude. Although the dataset is presented as a major contribution, specific inter-annotator agreement statistics are still missing.
The rebuttal also includes additional experiments involving ZNSM-iDB, RT-DETRv4-S, and YOLOv26n. These experiments go beyond what is permitted under the MICCAI rebuttal guidelines and should only be considered as background reference.
Review #3
- Please describe the contribution of the paper
This paper introduces Det-Y: a multicentered dataset that contains 5,444 annotated Mycobacterium tuberculosis (MTB) bacilli from 1,970 sputum smear images and a new framework, YOLO-Y: built on YOLOv11.The proposed YOLO-Y contains an adaptive ROI strip block, which is an attention module placed in the detection head. This module is used to capture the slender, rod-shaped morphology of Mycobacterium tuberculosis (MTB) bacilli.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
a) TB diagnosis is a clinically motivated and globally relevant problem. b) DET-Y dataset is a multicentered dataset that was taken with different microscopes at different resolutions, making this dataset suitable for a clinical setting. c) Adaptive RoiStrip Block is conceptually sound and uses asymmetric convolutions, which can better capture elongated, rod-shaped MTB bacilli as demonstrated by results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
a) Potential data leakage: If the overlapping patches of the same image fall in both the train and test splits, the data leakage will occur, which can directly affect the evaluation parameters. Authors need to clarify this. b) The model-level detection of individual bacilli does not translate to accuracy at the patient level. Therefore, the clinical utility of the proposed method is somewhat unclear. c) The size of Det-Y and choice of hyperparameters may lead to possible overfitting, and reported results might be overly optimistic.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
a) The dataset is derived from existing datasets. The total annotated images in the parent datasets and the total annotated images in the Det-Y dataset should be mentioned so that the contribution can be justified. b) The proposed model should also be evaluated on publicly available MTB bacilli detection datasets for robustness. c) Splitting criteria of the dataset should be clearly explained, i.e., the dataset was split at the patient level or image level. d) The loss curves should be calculated to better observe the training and validation phase. e) Det-Y comes from only two data centers; more data can be added to cover the diversity in medical equipment and geography. f) YOLOv10n, YOLOv9s indicate nano and small scales of the respective models. The scale of the used YOLOv11 backbone should be mentioned. g) The learning rate of 0.6 is quite large. The reason for choosing an unusual learning rate should be mentioned. h) The model was trained for 2000 epochs which can create the problem of overfitting. i) In Table 1, although the precision of YOLO-Y is better than YOLOv11, YOLOv9s outperforms YOLO-Y in the case of precision; this should be explained.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
a) This paper introduces Det-Y: a multicentered dataset that contains 5,444 annotated Mycobacterium tuberculosis (MTB) bacilli from 1,970 sputum smear images and a new framework, YOLO-Y, built on YOLOv11. b) The paper solves a clinically motivated problem using a novel architecture that can be considered a contribution.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the clarification in their rebuttal regarding the potential data leakage, and keeping in view their dataset contribution, I would be happy to see the paper accepted.
Author Feedback
We thank the reviewers for their constructive feedback. To address concerns about novelty, we emphasize our dual contributions: a meticulously re-annotated multi-center benchmark Det-Y, and a tailored lightweight detector YOLO-Y. Both are valuable to the Tuberculosis (TB) detection community. Below we respond to all major concerns.
1.Data leakage (R1, R2, R3) We apologize for the confusion. We have clarified in the revised manuscript that splits were always performed at the original image level (by patient ID) before any cropping; overlapping patches are generated independently within each split. This guarantees zero information leakage.
2.Generalization claim (R1, R2) We agree “strong generalization to Center-2” was overstated and have removed it. Center-1 (Ziehl-Neelsen) and Center-2 (Auramine-O) use different staining; direct cross-center testing is clinically unrealistic. Instead, we conducted cross-dataset evaluation on external Ziehl‑Neelsen Sputum Smear Microscopy Image Database (ZNSM‑iDB) (same staining as Center-1). Trained on Center-1, tested on ZNSM-iDB: YOLO-Y achieves mAP@0.5=0.607, F1=0.599, outperforming YOLOv12 (0.587, 0.595) and YOLOv11n (0.557, 0.574). This demonstrates genuine generalization under identical staining.
3.Dataset contribution – our primary contribution (R2, R3) The original public datasets had no bounding boxes for individual bacilli. Our pathologists manually annotated 5,444 bacilli across 1,970 images from two centers with distinct staining protocols, enhancing real-world diversity. In the revised manuscript, we add information on annotation protocol and inter-annotator agreement for more comprehensive introduction. A subset is already released; full data upon acceptance. This dataset alone fills a critical gap and will enable future research beyond our model.
4.Method novelty and statistical significance (R2) We agree the Adaptive RoiStrip Block is a task-driven engineering adaptation (first use of hierarchical strip convolutions for TB detection with level‑adaptive kernels 7/5/3). We have toned down overclaims. To address small gains: we repeated all main experiments 5 times (paired t‑tests, normality verified). On Center-1, YOLO-Y achieves mean mAP@0.5:0.95 of 0.573±0.002 vs. YOLOv11n’s 0.571±0.003 (p=0.03). On Center-2, 0.728±0.003 vs. 0.724±0.004 (p=0.01). These statistically significant gains mean fewer missed bacilli in screening—a clinically valuable outcome.
5.Efficiency metrics (R2, R3) We now report Params, GFLOPs, and latency for all models in Table 1.YOLO-Y: 3.0M params, 7.8 GFLOPs, 1.8–3.0 ms latency, comparable to YOLOv11n (2.6M, 6.3 GFLOPs, 1.8–3.4 ms) and far more efficient than YOLOv9s (7.2M, 26.7 GFLOPs, 3.7–4.9 ms). This supports the “lightweight” claim.
6.Non‑YOLO and new-YOLO comparisons (R2, R3) We trained RT‑DETRv4‑S (10.2M params) and YOLOv26n (2.4M params) on Center-1 and tested on ZNSM-iDB. Both underperform YOLO-Y (mAP@0.5: 0.425 and 0.522 vs. 0.607).
7.Patient‑level utility (R3) We added a discussion: a simple threshold (“positive if ≥1 bacillus”) converts outputs to patient‑level decisions; we will release a script. This is noted as future work.
8.Training hyperparameters and overfitting (R2, R3) LR=0.6 chosen via grid search (0.1–0.9). 2000 epochs with cosine decay and early stopping (patience=100). Loss curves show convergence by ~500 epochs with no overfitting.
9.Writing and formatting (R1, R2, R3) Fixed typos (“evaluation metrics”, Eq. 3 description), 0.016 miscalculation, bibliography (Springer style, added PANet, updated YOLOv9/v10/v12), and hyperlink colors.
We believe these revisions substantially strengthen the paper and respectfully ask the Area Chair and reviewers for acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a multi-center TB smear bacilli detection benchmark paired with a practical lightweight YOLO adaptation. In rebuttal, please address potential data leakage from overlapping patch splits by clarifying. Substantiate cross-center generalization and strengthen dataset contribution. Broaden or justify baselines beyond YOLO and report proper efficiency metrics like parameters, FLOPs, and latency. Clarify model scales, hyperparameters, and fix writing and reference issues. Please respond carefully and comprehensively to all reviewer concerns and do not feel limited by this meta-review.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I agree with the majority of the reviewers’ recommendation, and rank this paper within the top 50% of the rebuttal-stage papers in my batch. The key contribution lies in the establishment of a multi-center database, although the value of the work is not limited to the dataset itself. In particular, the proposed Det-Y framework presents ideas that are worthwhile for discussion within the MICCAI community.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper proposes a multi-center TB smear bacilli detection benchmark and a lightweight YOLO adaptation. The dataset contribution is recognized by two reviewers and has potential community value.
However, two critical issues prevent acceptance. First, Reviewer #2 notes that the authors introduced new experimental results in the rebuttal, which violates the MICCAI rebuttal guidelines prohibiting new experiments. Such a procedural violation is grounds for rejection and undermines trust in the evaluation. Second, the methodological contribution of the Adaptive RoiStrip Block is marginal, with 5-run statistics showing overlapping confidence intervals against the baseline, providing insufficient evidence of meaningful improvement.
While the dataset is a contribution worth sharing with the community, the combination of a procedural violation and a weakly validated method does not meet the acceptance threshold. I recommend reject. The authors are encouraged to resubmit to a future venue with a properly revised manuscript that includes the cross-dataset evaluation and stronger baselines in the main text from the outset, and to report inter-annotator agreement and proper efficiency metrics to fully substantiate the dataset and model contributions.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Regardless of new experiment introduced, the paper’s main contribution is data set and clarifications make the paper’s contribution more clear.