List of Papers Browse by Subject Areas Author List
Abstract
While deep learning models excel at medical segmentation, they typically do not leverage structured clinical descriptors that could inform feature learning. Here we show that integrating structured clinical descriptors into segmentation models guides learning and boosts performance. We introduce BioGuide, a lightweight multi-task learning framework that integrates fine-grained radiological descriptors as auxiliary targets to guide feature learning.
A small MLP head is attached to the encoder bottleneck of standard segmentation networks, encouraging latent representations to align with clinically meaningful patterns, boosting potential interpretability. We evaluate BioGuide across four diverse 2D and 3D brain MRI datasets for Multiple Sclerosis (MS) and Focal Cortical Dysplasia (FCD) using state-of-the-art segmentation models. Our results show that incorporating biomedical guidance consistently improves segmentation scores. BioGuide yields a consistent improvement of up to 3.1\% in Dice score and 8.5\% in lesion detection rate across all tested datasets, demonstrating superior data efficiency. Code is available at https://github.com/NadezhdaAlsahanova/BioGuide.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5935_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/NadezhdaAlsahanova/BioGuide
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AlsNad_BioGuide_MICCAI2026,
author = { Alsahanova, Nadezhda AND Statsenko, Yauhen AND Bernstein, Alexander AND Hamoudi, Rifat AND Sharaev, Maxim},
title = { { BioGuide: Biomedically-Guided Segmentation for Medical Tasks } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
BioGuide is a lightweight, multi-task learning (MTL) framework. The goal is to feed clinical domain knowledge into deep learning models to guide their feature learning.
- 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.BioGuide adds a tiny MLP to the encoder bottleneck of standard networks like nnU-Net or Transformers. The plug-and-play design is lightweight.
2.BioGuide trained on 70% to 80% of the data actually beat baseline models using 100%.
3.This work shows interpretability. PCA results show that the model’s feature space actually forms clusters based on real medical signs, like blurred borders or specific anatomy.
- 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.If a patch doesn’t have a lesion, the authors set the entire guidance vector to all zeros. However, this vector includes Temporal Lobe. If a model sees a healthy Temporal Lobe, it gets penalized for predicting “Temporal Lobe” because there’s no lesion.
2.The authors attached the MLP head to the encoder bottleneck, but the bottleneck has the lowest resolution. Asking the model to predict blurred boundaries (tiny textures) at the exact layer where those textures have been compressed away is physically illogical.
3.The comparison between BioGuide and nnUNet is unfair. The authors give the model the all-zero labeling, it’s basically being told “is there a lesion here?” The authors shoud improve the baseline, like also give the “lesion: yes or no” head to nnUNet
- 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.
(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?
method novelty
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors have developed a lightweight multi-task learning framework (BioGuide) where predicting radiological descriptors is used as the auxiliary task to the main segmentation task, using an MLP head added to the encoder bottleneck layer. They argue that the auxiliary target encourages latent representations to align with clinically meaningful patterns. The frameworks is generic and they evaluate BioGuide on top of state-of-the-art segmentation models using two brain MRI datasets in Multiple Sclerosis (MS) and Focal Cortical Dysplasia (FCD).
- 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 developed framework is light and generic enabling it to be added easily to SOTA segmentation models. The framework is effective in improving segmentation performance in the two segmentation tasks in MS and FCD.
- 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 proposed framework addresses a joint segmentation and classification/regression task, a configuration already well-explored in prior literature [1-2]. Consequently, the methodological novelty appears limited. Furthermore, the clustering by folds observed in Figure 3 suggests potential model overfitting; while the radiological descriptor task may provide some regularization, its impact is unclear. Notably, the PCA visualization fails to show any meaningful clustering or underlying patterns, further questioning the robustness of the learned representations.
[1] Huang L, Shao Y, Yang H, Guo C, Wang Y, Zhao Z and Gong Y (2024) A joint model for lesion segmentation and classification of MS and NMOSD. Front. Neurosci. 18:1351387.doi: 10.3389/fnins.2024.1351387 [2] Ryu, H., Shin, S.Y., Lee, J.Y. et al. Joint segmentation and classification of hepatic lesions in ultrasound images using deep learning. Eur Radiol 31, 8733–8742 (2021). https://doi.org/10.1007/s00330-021-07850-9
- 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?
Limited technical novelty. Though the dice improves by 2.5-3% and detection rate improves by 7-8% for the FCD task, the improvement is marginal for MS (1.1 for Dice & 2.1 for detection rate).
- 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.
Thanks to the authors for clarifying BioGuide’s contribution as a lightweight auxiliary module utilizing local radiological descriptors for segmentation. However, I agree with R1 that bottleneck placement captures global context but sacrifices fine local details. Performance mismatches in Tables 1 and 2 (restricted internal test sets and a large validation/test Dice gap on Internal Dataset 1) alongside Figure 3 suggest the method acts more as a regularizer against small-data overfitting than a robust, scalable methodological contribution.
Review #3
- Please describe the contribution of the paper
This paper proposed BioGuide, which is a lightweight multitask framework that enhances medical image segmentation by incorporating structured clinical descriptors as auxiliary supervision, guiding feature learning toward clinically meaningful patterns. The proposed method and auxiliary tasks are effective. This approach consistently improves segmentation performance across multiple brain MRI datasets.
- 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 paper proposes a new method BioGuide, which is an interesting and novel idea. By introducing constraints in the latent space, the model learns more informative feature representations, leading to improved MRI segmentation performance.
- 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.
Since different types of brain lesions may exhibit distinct clustering patterns, providing such visualizations across all datasets would offer more comprehensive insight.
Minor: texts in figure 3 are small.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, the idea is interesting. The paper presents a compelling and well-motivated approach: incorporating latent space constraints (auxiliary tasks) enables the model to learn more meaningful feature representations, which in turn leads to improved MRI segmentation performance. Both qualitative and quantitative results demonstrate the effectiveness of BioGuide.
- 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.
Method novelty with improved performance
Author Feedback
We thank the Meta-Reviewer and reviewers for constructive feedback. We address the central concern - methodological novelty - and the specific technical points.
1.Novelty vs. joint segmentation+classification (R1, R2, Meta). R2 cites Huang 2024 [1] and Ryu 2021 [2] as overlapping prior work. However, BioGuide differs in what the auxiliary task is, in how it is implemented, and in what it delivers for segmentation. (a) [1,2] use disease-level labels (MS vs. NMOSD; four hepatic lesions). BioGuide instead encodes fine-grained radiological signs - blurring, transmantle sign, lobar location - the precise descriptors radiologists use, tightly coupled to lesion morphology. To our knowledge, no prior work uses radiological signs as auxiliary targets. (b) Unlike [1,2], where classification is primary and segmentation gains are marginal, segmentation is critical for MS and FCD. [2]: JI rises only from 68.5±10.3 to 70.0±10.9 with benign/malignant labels and does not improve with 4-class labels (68.5±12.2). [1]: joint-model precision is below baseline (73.9 vs. 74.1 for UNet); no small-cluster detection is reported, and Fig. 10 shows only refinement of already-detected large clusters. For dispersed pathologies like MS, finding all clusters matters more than refining contours. BioGuide improves every metric, including detection. (c) Architecture: [1] uses a dual-branch backbone with a Lesion Guidance Module; [2] requires user clicks. BioGuide is a single MLP at the bottleneck - no backbone change, no user input - validated on two SOTA backbones, four datasets. Knowledge-blind ablations (Random/Shuffled) isolate medical semantics from architectural regularization and are absent from both cited works.
2.Zero-vector for healthy patches (R1.1). “Temporal Lobe” is a binary property of a lesion, not the patch’s location. It describes lesion location, which directly influences morphology (e.g., temporal lesions are typically larger than occipital ones). All-zero labels for healthy patches are equivalent to the standard “no-object” target in multi-label detection.
3.Bottleneck placement (R1.2). Lesion characteristic vectors are provided as binary labels at the patient level, therefore, the precise spatial coordinates of “Blurring” are unknown. The model is not asked to localize “Blurring” voxel-wise, only to determine whether it exists in the patch. The bottleneck captures the highest-level, spatially compressed semantic context, ideal for this task. Backpropagating the patch-level loss through the encoder forces early and mid-level filters to become sensitive to clinically relevant signs.
4.”Fair” baseline with a yes/no head (R1.3). This is exactly what Random Noise and Shuffled Medical Labels test: identically-sized MLP heads, jointly trained, isolating signal content from the act of adding a head. A “lesion yes/no” head is a degenerate case of shuffled-label control. A shuffled-label vector containing at least one non-zero entry still provides the model with a global signal that a lesion exists within the patch.
5.PCA (R2,R3). Fold clustering in the baseline indicates the model memorizes fold-specific intensity distributions - poor generalization. BioGuide breaks this and reorganizes the latent space by radiological signs, the desired outcome. Better generalization is confirmed quantitatively: in Table 2, BioGuide trained on 70% data outperforms the baseline trained on 100% - data efficiency, not overfitting. Due to the 8-page limit, additional PCA figures could not be included but will be made available on GitHub upon acceptance.
6.”Marginal” MS gains (R2). MS gains (Dice +1.1%, detection +2.1%) are modest but consistent across all folds. MS lesions are higher-contrast on FLAIR and the nnU-Net baseline is already near-optimal. For FCD, where lesions are subtle and heterogeneous, gains are larger, precisely because clinical descriptors carry more discriminative information there.
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 manuscript had very mixed reviews. The authors should focus on addressing the limited novelty issues raised by R2 and R1 in the rebuttal stage.
- 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.
Authors present a light weight MLP layer to help regularise model performance trained on auxiliary tasks related to image quality.
Overall the biggest weakness was the limited novelty, essentially a light weight layer added to a network to help regularise weights via an auxiliary tasks. There were also concerns about the fairness of the experiments especially with the labels for the tasks.
The authors did a good job at addressing most critiques. The main remaining point after the rebuttal is the interpretation of what the MLP layers are doing and how they are helping improve model performance. I would suggest the authors consider alternative explanations for the observed model improvements and consider modulating their claims to reflect what is demonstrated in the experiments.
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.
The paper proposes a framework for guided feature learning based on a multi-task learning scheme. The paper received mixed reviews before rebuttal and opinions are still divided after rebuttal. Since the paper positioned as a translation paper according to the submission by the authors, I agree with reviewer #2 and recommend reject due to limited novelty as the translation value/clinical impact beyond improved Dice scores / detection rates is not discussed / explored in the paper.
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.
This paper presents a lightweight and practical framework for incorporating structured radiological descriptors into medical image segmentation through auxiliary supervision. Reviewers agreed that the approach is simple, broadly applicable, and consistently improves segmentation performance across multiple datasets and architectures.
The main concerns relate to the degree of methodological novelty and the interpretation of experimental analyses. While the overall framework builds upon established multi-task learning principles, the rebuttal clarifies important distinctions from prior joint segmentation-classification approaches, particularly the use of fine-grained radiological descriptors rather than disease-level labels. The rebuttal also provides reasonable explanations regarding the auxiliary supervision design and control experiments. Overall, despite some remaining concerns about novelty and evaluation depth, the paper offers a useful and clinically motivated contribution with consistent empirical improvements and a lightweight design.