List of Papers Browse by Subject Areas Author List
Abstract
Currently, unsupervised domain adaptive strategies proposed to overcome domain shift, are handicapped by the requirement of large amount of target data. On the other hand medical imaging problems and datasets are often characterized not only by scarcity of labeled and unlabeled data but also class imbalance. Few-shot domain adaptive object detection (FSDAOD) addresses the challenge of adapting object detectors to target domains with limited labeled data. However, existing FSDAOD works struggle with randomly selected target domain images which might not represent the target distribution, resulting in overfitting and poor generalization. We propose a novel FSDAOD strategy for microscopic imaging to tackle high-class imbalance and localization errors due to foreground-background similarity. Our contributions include: a domain adaptive class balancing strategy for few shot scenario and label dependent cross domain feature alignment. Specifically, multi-layer instance-level inter and intra-domain feature alignment is performed by enhancing similarity between the instances of classes regardless of the domain and increasing dissimilarity between instances of different classes. In order to retain the features necessary for localizing and detecting minute texture variations in microscopic objects across the domain, the classification loss was applied at feature-map before the detection head. Extensive experimental results with competitive baselines indicate the effectiveness of our proposed approach, achieving state-of-the-art results on two public microscopic datasets, M5 [12] and Raabin-WBC [10]. Our method outperformed both datasets, increasing average mAP@50 by 8.3 points and 14.6 points, respectively. The project page is available here.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/4161_paper.pdf
SharedIt Link: https://rdcu.be/dY6fE
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72390-2_10
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/4161_supp.pdf
Link to the Code Repository
https://im.itu.edu.pk/few-shot-DAODMI/
Link to the Dataset(s)
https://github.com/intelligentMachines-ITU/LowCostMalariaDetection_CVPR_2022
BibTex
@InProceedings{Ina_FewShot_MICCAI2024,
author = { Inayat, Sumayya and Dilawar, Nimra and Sultani, Waqas and Ali, Mohsen},
title = { { Few-Shot Domain Adaptive Object Detection for Microscopic Images } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15012},
month = {October},
page = {98 -- 108}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a framework for FSDAOD in microscopy. It includes a CBCP stage for source augmentation, and I2DA for effective domain alignment. Experiments on two scenarios demonstrate its efficiency.
- Please list the main strengths of the paper; you should write about 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 CBCP method is a data-driven method which seems to be very suitable for microscopy. The framework shows good performance improvement over previous methods. The paper is well organized.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- The few-shot DAOD should be abbreviated as FSDAOD, according to [1], not FDAOD. Please change it. [1] AsyFOD: An Asymmetric Adaptation Paradigm for Few-Shot Domain Adaptive Object Detection
- The proposed method is very complex. It combines various fragmentary components into a framework and many of them may just bring a little help for improving the performance (however, the authors do not give detailed ablation experiments). It is more preferable to design a neat framework with some most effective components.
- I think the proposed CBCP part can be applied to any general domain FSDAOD methods. It is just specially designed for microscopy. So let’s say on Raabin-WBC, I2DA do not show superiority over AcroFOD, why don’t we just combine CBCP with AcroFOD? What is the advantage of I2DA compared to existing FSDAOD approaches in this scenario? More in-depth analysis is required.
- Moreover, what is the advantage of CBCP over existing data augmentation methods? No comparison is given in this paper.
- 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 submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
Since the method is complex, it is recommended to open source the code for reproducibility.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
See weaknesses
- 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
Weak Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
See weaknesses
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Most of my concerns are addressed, so I raise my score to WA. Still, I suggest the authors to revise their method to make it neat, removing unnecessary components, and preserve those that are really useful.
Review #2
- Please describe the contribution of the paper
The authors have proposed to tackle the problem of few-shot domain adaptation for object detection in microscopy images. They do so by incorporating a copy-paste mechanism between source and target samples and the addition of the two new losses that compute the similarity and dissimilarity among the samples. They achieve significant performance gains compared to the relevant approaches.
- Please list the main strengths of the paper; you should write about 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 proposed mechanism works in two steps: ○ Copy-paste handles mixing target cells within the source data by incorporating gaussian and color smoothing to modulate natural blending. ■ This essentially handles the few-shot nature of the target domain as it is generalizing towards it. ○ Maintaining inter-domain alignment and intra-domain class consistency by employing cosine similarity/distance-based losses to capture similarity between instances of similar classes and dissimilarity between instances of dissimilar classes. Additionally, they add a generic classification loss to aid the adaptation process.
● The approach is sound and clearly explained.
● They outperform relevant approaches.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
● The authors have discussed the few-shot strategy in the implementation of their work; however, it would be interesting to see how quickly the algorithm adapts to the current target domain. In some instances, time can also be a constraint.
● As the distribution of target samples is across the source dataset, one can imagine the strategy being dependent on the source dataset. Therefore, would this algorithm work if the distribution shifts between the source and target data are significantly high? For example, bacteria phase and bacteria fluorescence usually have inverted values in which case the copy-paste might be difficult to employ as the model will shift towards source distribution rather than the target distribution.
- 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
Please refer to the weakness section, the concerns are explained in detail with the comments. I would like the authors to describe the concern for distribution shifts, which would help the readers understand the strengths and limitations of this work.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors have proposed a logical solution to few-shot domain adaptation for object detection in microscopy images by incorporating a set of similarity and distance based losses and a sound augmentation strategy. Even though concerns are related to the scope of the augmentation strategy, the approach could be widely adopted for few-shot learning; therefore, I am tending towards weak acceptance for now.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
The Authors have mostly addressed the concerns, and I am conforming the initial rating.
Review #3
- Please describe the contribution of the paper
This paper focus on the few-shot domain adaptive object detection probelm for microscopic images. An domain-generalized class balancing cut-paste strategy (CBCP) and Intra-Inter-Domain Feature Alignment technique (I2DA) are proposed to tackle the high-class imbalance and fg-bg similarity problem, respectively. The experiments on two different medical datasets (Malaria and Raabin-WBC) show superior performance than other methods.
- Please list the main strengths of the paper; you should write about 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.
- Few-shot domain adaptive object detection is rarely explored for medical imaging.
- The class balanced cut-paste strategy seems to be novel for extreme class imbalancing scenarios.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
Major concerns:
- The effectiveness of each part of the I2DA (e.g., L_sim, L_dis, and L_cls) module lacks validation.
- If other competative methods adopt the same YOLOv5 as base model? Minor concerns:
- In Section “Implementation Details”, many punctuation marks are missing.
- While the batch size is set to 4, 2% target data (only 0.08) for each batch seems confusing.
- Appendix exceeds the page limit (only 2 pages).
- 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 submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
- Please supplement the few-shot results of only CBCP module.
- Add the ablation study of each component of I2DA, and explain why set such a low threshold (e.g., 0.005 and 0.001) for each loss.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a relevant and somewhat novel contribution to the FDAOD task in medical images, but the innovation is somewhat incremental.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Most of my concerns are addressed.
Author Feedback
We greatly appreciate the reviewers’ insightful comments and recognition of our work’s strengths including the soundness and clarity of our approach (R1), the suitability of the CBCP method for microscopy (R3), and the novelty of addressing extreme class imbalance in medical imaging (R4). Below are our answers:
RWBC: Raabin-WBC
1- Data (R1) and code sharing (R3, R4): We will release the code upon acceptance, the data is already publicly available.
2- Effectiveness of each component of I2DA (R3, R4): | Ldis | Lsim | Lcls | Ldis+Lsim | Ldis+Lcls | Lsim+Lcls | Lsim+Ldis+Lcls M5 | 38.0 | 39.7 | 41.9 | 42.2 | 43.4 | 43.8 | 44.2 This validates the effectiveness of each component of the I2DA loss. Lsim+Ldis+Lcls results in improved results.
3- Base architecture of other competitive methods (R4): For a fair comparison, we have used YOLOv5, as used by AcroFOD and AsyFOD.
4- Few-shot results with only CBCP (R4, R3): Employing only CBCP (w/o I2DA), our results are still better than the source-only model: M5:39.7 vs 19.9 (Tab.1), RWBC:66.5 vs 27.2 (Tab.2). CBCP effectively handles class imbalance by strategically augmenting source and target cells based on target-domain visuals, leading to these improved results.
5- Performance of CBCP in extreme domain shifts (R1): Thank you for an interesting question, the shifts highlighted by R1 are quite challenging. However, input adaptation methods could partially reduce such domain shifts. Our algorithm could be run alongside such algorithms to overcome the limitations (especially class imbalance in src. & target) of input adaptation methods.
6- Effectiveness of our alignment (I2DA) over existing work (R3): AcroFOD alignment results w/o augmentation (M5:37.8 | RWBC:58.7) are much less than our alignment-only I2DA results(M5:44.2 | RWBC:66.5). AcroFOD alignment prioritizes target-similar examples, but we believe that small few-shot sets can’t fully represent the whole target population. Our approach extracts moderate knowledge from the target set, ensuring generalizability to larger test sets while optimizing performance across majority few-shot scenarios. 7- Effectiveness of CBCP over augmentation techniques and combination of AcroFOD and CBCP (R3): AcroFOD{align} and CBCP results in lower mAP (M5:45.8 and Raabin-WBC: 69.8) as compared to our method. Similarly, using our I2DA{align} along with AcroFOD aug. also results in decreased mAP (M5:43.0 and Raabin-WBC: 67.9). NOTE: our method (CBCP+I2DA) results are M5:48.9(Tab.1), Raabin-WBC:70.7(Tab.2)
8- I2DA results on RWBC (Table:2) (R3): As shown above (6&7), I2DA and CBCP complement each other. Please note even using alone I2DA, gives better results than SOTA on 3 & 4 shots on RWBC and all shots on other experiments.
9- How quickly the algorithm adopts (R1): We train for 100 epochs and at around 80-85 epochs it starts generalizing to a larger test set.
10- 2% of batch size (R4): Thank you for pointing this out. We will make it more clear in the final manuscript. We used a batch size of 4, in our experiment, due to limited GPU. However, training on larger batch sizes, it is suggested that the real target samples from the few-shot set should comprise 2% of the total batch for better generalization.
11- Why so low thresholds (R4): We chose low Lambda 1,2,3 weights to scale the complete I2DA loss with the YOLOv5 detection loss.
12- Appendix and typos (R3, R4): We will update it in the final version.
Meta-Review
Meta-review #1
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
N/A
Meta-review #2
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper proposes a few-shot domain adaptation method for cell detection in microscopy images. The authors achieve this by incorporating a copy-paste mechanism between source and target samples and adding two new losses that compute the similarity and dissimilarity among the samples. The reviewers highlighted the well-designed method for cell images, its good performance, and its organization as strengths. However, they expressed concerns about the detailed analysis of the adaptation process of the proposed method and the effectiveness of each module. The rebuttal addressed these issues, and after considering it, all reviewers agreed to accept the paper (WA, WA, WA). The meta-reviewer recommends accepting it.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
This paper proposes a few-shot domain adaptation method for cell detection in microscopy images. The authors achieve this by incorporating a copy-paste mechanism between source and target samples and adding two new losses that compute the similarity and dissimilarity among the samples. The reviewers highlighted the well-designed method for cell images, its good performance, and its organization as strengths. However, they expressed concerns about the detailed analysis of the adaptation process of the proposed method and the effectiveness of each module. The rebuttal addressed these issues, and after considering it, all reviewers agreed to accept the paper (WA, WA, WA). The meta-reviewer recommends accepting it.