List of Papers Browse by Subject Areas Author List
Abstract
Microscopic analysis of bone marrow aspirate smears is a critical routine procedure for the initial screening and diagnosis of multiple myeloma (MM). However, manual interpretation is subjective and labor-intensive, further complicated by the fine-grained morphological heterogeneity and inter-class similarities of bone marrow cells. While vision foundation models like DINOv3 have revolutionized natural image understanding, they lack the domain-specific inductive bias required to capture subtle pathological textures in medical cytology. To address these challenges, we first construct a comprehensive, expert-annotated MM diagnosis dataset encompassing six fine-grained cell categories to reflect real-world clinical scenarios. We then propose SF-DINO, a spatial-frequency adapted foundation model that efficiently adapts DINOv3 for MM diagnosis. Specifically, we introduce a parallel spatial-morphology (SM) adapter to inject high-frequency local morphological cues into the backbone without disrupting pre-trained semantics. Additionally, we devise a frequency-selective hashing attention (FSHA) module to model global long-range dependencies efficiently via spectral clustering. Extensive experiments demonstrate that SF-DINO significantly outperforms six state-of-the-art methods, showing impressive generalization and establishing a new benchmark for automated, fine-grained MM diagnosis. Our code is available at https://github.com/yzygit1230/SF-DINO.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0787_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yzygit1230/SF-DINO
Link to the Dataset(s)
https://github.com/yzygit1230/SF-DINO
BibTex
@InProceedings{YeZha_SFDINO_MICCAI2026,
author = { Ye, Zhaoyi AND Shen, Hui AND Zhou, Fuling AND Hu, Xinglong AND Wang, Du AND Mei, Liye AND Lei, Cheng},
title = { { SF-DINO: Spatial-Frequency Adapted Foundation Model for Multiple Myeloma Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
- The authors assemble a BM smear data set from multiple myeloid patients with thousand of annotated single white blood cells. They design and evaluate a novel FFT based method that seems to outperform SOTA methods.
- 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.
- Good intro, relevant question
- compilation of a annotated MM dataset
- interesting FFT based method
- improvement over other methods
- evaluation on external data
- 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.
- The paper works on single cell images but not show how the improved classification accuracy translate to diagnosing patients.
- A statistical analysis is missing: Are differences statistically significant?
- Authors should apply their method to other published BM smear data, like the one from ref 12
- Not clear if SF DINO is retrained on Acevedo or zero-shotted.
- 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.
- 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
- Abstract:
- I suggest so specify ‘frequency’ as is could relate to image properties or cell type frequency
- Avoid ‘impressive’
- Introduction
- I wouldn’t call challenges ‘dilemmas’
- M&M
- page 3 typo: ‘them. N’
- ‘we acquire 400 bone marrow cell images’ i guess you want to say ‘we acquire 400 bone marrow images’? yolo comes later…
- 2.1: provide the number of patients. Do you also have controls?
- please elaborate on what spectral clustering in the frequency domain does.
- can you provide evidence for ‘its global self-attention mechanism inherently lacks the inductive bias to capture high-frequency morphological details essential for fine-grained cell differentiation’
- Experiments
- typo: plasm cells
- when do you think an AUC is ‘impressive’?
- provide error bars to all your metrics
- a 2% increase in in recall is nice, but a 6% variance is a bit worrying.
- the macro-averaged metrics may be masking poor performance on minority classes
- Fig 2e not discussed: What do you make of the GradCAM images?
- If your variance is 2.32, then reporting two digits behind the comma (87.50) makes no sense. 88±2% would be appropriate.
- Is SF DINO retrained on Acevedo? or zero-shotted?
- 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?
- relevant topic, well described methodology
- limited external validation though
- some statistical issues
- 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.
Solid answers to the reviewer’s point, willing to adapt wording in revised version.
Review #2
- Please describe the contribution of the paper
The paper presents SF-DINO, a spatial-frequency adapted foundation model designed for the fine-grained classification of bone marrow cells for Multiple Myeloma (MM) diagnosis. The main contributions include: 1.Building a comprehensive, expert-annotated single-cell dataset for MM diagnosis comprising 6,961 images across 6 fine-grained categories, reflecting real-world clinical complexity. 2.Introducing a parameter-efficient Spatial-Morphology (SM) adapter into the DINOv3 backbone using depth-wise convolutions to capture high-frequency local morphological textures. 3.Proposing a Frequency-Selective Hashing Attention (FSHA) module that operates in the spectral domain (via FFT) to model global structural dependencies and periodic textural patterns efficiently through hash-based clustering. 4.Achieving competitive performance on the internal dataset against recent state-of-the-art architectures and demonstrating strong cross-domain transferability on the public Acevedo dataset.
- 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 methodology addresses a critical bottleneck in deploying vision foundation models (like DINOv3) to cytopathology: the lack of domain-specific inductive bias. The parallel incorporation of the SM adapter allows the network to inject necessary local spatial cues without suffering from catastrophic forgetting of the pre-trained semantics. 1.Clinically Meaningful Task Formulation: Moving beyond simplified binary classification (normal vs. myeloma), the authors tackle a 6-class fine-grained categorization that accounts for morphological confounders like erythroid and granulocytic lineages. This is highly relevant for real-world bone marrow microenvironments. 2.Rigorous Baseline Comparisons: The experimental section is comprehensive, benchmarking against a strong and highly relevant set of recent state-of-the-art architectures. The inclusion of pure CNNs, Transformer-hybrids (InceptionNeXt), and recent state space models (MedMamba, Swin-UMamba) provides a very convincing baseline for evaluating the classification capability of SF-DINO. 3.Strong Generalization: The cross-domain evaluation on the external Acevedo dataset (achieving 95.35% accuracy) strongly supports the claim that the extracted spatial-frequency features are robust against variations in staining and image acquisition.
- 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.Missing Clinical Aggregation (Slide/Patient Level): While the model achieves high accuracy on single-cell crops, clinical MM diagnosis ultimately relies on the proportion of plasma cells at the whole-slide or patient level. The paper lacks an analysis of how the single-cell classification accuracy (89.95%) translates to patient-level diagnostic confidence or cell counting accuracy, which is the actual clinical endpoint. 2.Lack of Hyperparameter Sensitivity Analysis: The FSHA module introduces several key components, notably the number of hash rounds ($n_{hash}$) and the number of buckets ($Q/K$). There is no ablation study exploring how sensitive the model’s accuracy and inference speed are to these specific hyperparameters. 3.Baseline Implementation Details: Although the comparison with models like MedMamba and InceptionNeXt is a strength, the paper lacks details on how these baselines were implemented. Were they trained from scratch, or were pre-trained weights (e.g., ImageNet) utilized? 4.Please correct the typo in the heading of Section 3.4 (“Abaltion Analysis” should be “Ablation Analysis”.
- 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
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.
(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?
I recommend a “Weak Accept” for this submission. The paper proposes a well-designed, parameter-efficient framework (SF-DINO) that successfully adapts a natural-image vision foundation model to the highly specialized domain of hematopathology. The dual mechanism—using depth-wise convolutions for local morphology and spectral clustering for global dependencies—is theoretically sound and empirically validated. Furthermore, the dataset construction and the choice of advanced baseline architectures (including recent state space and hybrid models) reflect a high-quality experimental design. However, the novel FSHA module lacks a hyperparameter sensitivity analysis. While the cell-level metrics are strong, the manuscript stops short of demonstrating how these individual cell predictions aggregate into a patient-level clinical diagnosi, which slightly weakens the clinical impact of the application study.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The authors have fully responded to all my concerns. They supplemented hyperparameter sensitivity analysis of the FSHA module, clarified the pre-training initialization strategy of all baseline models, and reasonably clarified the research scope of single-cell morphological identification and distinguished it from patient-level clinical definitive diagnosis, and clearly arranged patient-level aggregation analysis as future work. The typo error is easy to revise. All technical deficiencies in the original manuscript have been well addressed.
Review #3
- Please describe the contribution of the paper
SF-DINO adapts DINOv3 for fine-grained bone marrow cell classification using a spatial-morphology adapter and a frequency-selective hashing attention module.
- 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.
1.Fine-grained bone marrow cell recognition is directly related to MM cytology workflows. 2.The six-class setting is more useful than a binary MM-versus-normal task. 3.The method is easy to follow. 4.The experiments include ablation, speed comparison, Grad-CAM visualization, and an external dataset.
- 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.Patient counts, smear counts, fold construction, class distribution per fold, DINOv3 checkpoint, hyperparameters, and FSHA implementation details are missing. 2.The experiments classify cropped single-cell images. MM diagnosis is patient-level and depends on plasma cell proportion, smear context, clinical information, and often immunophenotyping. The paper should frame the task as cell classification for assisting MM diagnosis. 3.SF-DINO reports 89.95 accuracy and 87.13 F1.MedMamba has 89.37 accuracy, and Swin-UMamba has 86.10 F1.The margin is small relative to the standard deviations. 4.The SM adapter is close to standard convolutional adapter tuning, and FSHA combines FFT processing with hash attention. Simpler alternatives such as LoRA, partial DINO fine-tuning, plain convolutional adapters, or frequency attention without hashing are missing.
- 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 does not provide sufficient information for reproducibility.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The problem is worth studying, but the evidence is not enough. The key issues are overstatement of single-cell classification as MM diagnosis, small gains relative to variance, and missing simple adaptation baselines.
- 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.
This paper is a borderline paper. The rebuttal clarifies more implementation details, but some main concerns remain: the dataset contains only 40 patients, and the performance gains over baselines are modest with overlapping standard deviations. If cells from the same patient are in both training and test folds, their performance may be inflated by information leakage.
Author Feedback
We thank all reviewers for their insightful and constructive feedback. We are encouraged that they recognized the value of our MM diagnosis dataset and SF-DINO method (Meta, R1-R3). We provide responses to the remaining concerns below. 1.Clinical Translation (Meta, R1-R3) We agree that definitive MM diagnosis is a patient-level clinical decision and cannot be made solely from cropped single-cell images. In clinical practice, MM diagnosis relies on the quantification and morphological assessment of monoclonal plasma cells, together with smear-level context, clinical information, immunophenotyping, and other ancillary tests. Therefore, in this work, “diagnosis” should be interpreted as morphology-based diagnostic assessment rather than standalone definitive patient-level diagnosis. SF-DINO focuses on the accurate, fine-grained identification of six critical bone marrow cell lineages. By automating this process, it provides the essential morphological evidence required for reliable plasma-cell quantification, thereby robustly assisting MM diagnosis. We will make this scope explicit in the final version and discuss patient-level aggregation as future work. 2.Novelty (Meta, R3) Clinically, we introduce a fine-grained dataset encompassing malignant monoclonal plasma cells alongside 5 common cell lineages (erythroid, granulocytic, lymphoid, monocytic, and other cells). Methodologically, SF-DINO provides a parameter-efficient alternative to fine-tuning or generic LoRA. Unlike standard convolutional adapters, our SM-Adapter is specifically designed to inject high-frequency local morphological cues crucial for cytopathology. Furthermore, the FSHA module utilizes frequency-domain hash attention to efficiently filter background noise and model global periodic chromatin textures with minimal computational overhead. As shown in Table 2, SF-DINO significantly boosts performance while adding negligible trainable parameters (1.16M). 3.Dataset Details (Meta, R1, R3) Our MM diagnosis dataset originated from 40 patients (without a control group). We employed a stratified 5-fold cross-validation, strictly preserving the class distribution (detailed in Sec 2.1) across all folds. For the Acevedo dataset, we retrained all models to evaluate cross-domain transferability. Table 3 shows the white blood cell classification results. Regarding [12], their dataset focuses on regular bone marrow smears without oil immersion, presenting a domain gap. Furthermore, their reported recall for plasma cells is 84%, which is lower than our MP cell recall value, underscoring our superiority in identifying critical myeloma cells. 4.Implementation Details (R1-R3) We trained all six comparative methods initialized with ImageNet pre-trained weights, and followed the original parameter configurations in their respective 2024-2025 publications. We used the frozen DINOv3 ViT-B checkpoint. SM-Adapter uses a 64-dim bottleneck. FSHA used 2 hash rounds, a chunk size of 112, and a channel reduction of 4.We implemented FSHA via 2D Real FFT, applying hash-based sparse attention, and restoring via IFFT. Spectral clustering refers to using a hashing routing branch for these frequency features. By hashing similar spectral components into identical buckets, it explicitly groups dominant periodic textural patterns. 5.Statistical & Sensitivity Analysis (Meta, R1, R2) Using the same 5-fold, paired t-tests between SF-DINO and six methods demonstrated that SF-DINO achieved statistically significant improvements in recall and F1-score metrics (p < 0.05). Our sensitivity analysis shows a clear trade-off: increasing n_hash and Q/K chunk size improves diagnostic accuracy but naturally decreases inference speed (dropping by ~1.2 img/s per additional hash round). Results show that setting n_hash=2 and chunk size=112 provides the optimal balance, securing peak diagnostic performance while maintaining a highly competitive inference speed of 22.69 img/s.
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 received mixed review scores of weak accept and weak reject. The major contributions and strengths summarized by reviewers include the dataset contribution, the FFT-based method, and good performance on the datasets. Reviewers also had a list of comments and concerns to address. Some major ones include: the translation from the single cell classification to clinical diagnose is not connected which makes the paper title and problem over-claimed; the current experiment section is lack of statistical analysis, experiments on other public datasets, and sensitivity analysis; the justification on small improvement margin; and the differences with related literatures; along with other comments from individual reviewers.
- 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.
Two reviewers’ concerns were solved after the rebuttal. The third reviewer still has some concerns. After reading the paper, rebuttal and the reviews after the rebuttal, the AC recommends accepting this paper which should be revised based on the three reviewers’ feedback.
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.
The scores remain mixed after the rebuttal. The three reviewers have recognized some novelty in the proposed frequency-domain method for adapting a foundation model to single-cell multiple myeloma images. The rebuttal clarified several implementation details, and included the sensitivity analysis answering to several of the reviewers answers . The remaining issues remain the patient-wise diagnosis which is left for future work and the dataset size (40 patients but with > 6k single-cell images). Considering the grounded medical motivation, the methodological novelty and the extensive comparisons showing some improvement, I support the acceptance of this paper despite its shortcomings.
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.
the paper addressed the majority of the concerns and 2 of 3 reviewers accepted the paper after rebuttal, I myself concurred with them too.