List of Papers Browse by Subject Areas Author List
Abstract
Focal Cortical Dysplasia (FCD) is a leading cause of drug-resistant focal epilepsy, and its identification on MRI can enable potentially curative surgical treatment. FCD features are often subtle and heterogeneous and current deep learning models for automated detection exhibit limited sensitivity. In clinical workflows, prior knowledge about the suspected localization of seizure onset can guide lesion search. We hypothesize that this prior knowledge can similarly guide a deep learning model in the presence of ambiguous imaging features. We propose a text-guided graph neural network that incorporates prior knowledge to guide FCD detection on cortical features derived from MRI. We further introduce a cluster-wise distance-based gating mechanism that leverages the per-vertex distance regression head at inference to suppress false-positive predictions. Experimental results demonstrate that hemisphere- and lobe-level localization information improves sensitivity and segmentation quality, while the gating mechanism substantially increases specificity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5707_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Wektor607/FCD-TextGuide
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MikGer_Text_MICCAI2026,
author = { Mikhelson, German AND Lange, Annalena AND Rüber, Theodor AND Working Group, MELD AND Schultz, Thomas},
title = { { Text Guidance and Distance Gating Improve Detection and Segmentation of Focal Cortical Dysplasia } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the challenging task of automated detection and segmentation of Focal Cortical Dysplasia (FCD) from MRI, a leading cause of drug-resistant focal epilepsy. The authors build upon the state-of-the-art MELD framework, which uses surface-based graph neural networks. The core novelty is the integration of textual prior knowledge about the suspected anatomical localization of the lesion (e. g. , hemisphere or lobe) to guide the model’s attention. The proposed framework employs a frozen vision encoder (MELD) and a frozen text encoder (RadBERT), with fusion achieved via a GuideDecoder architecture adapted for cortical surface meshes. Additionally, the authors introduce a cluster-wise distance-based gating mechanism at inference time, leveraging the existing distance regression head of MELD to suppress false-positive predictions, thereby significantly improving specificity. Experiments on the large, multi-center MELD dataset demonstrate that incorporating coarse localization cues improves sensitivity and segmentation IoU, while the gating mechanism effectively restores specificity compromised by the text-guided model.
- 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) The paper makes a compelling clinical argument: in presurgical epilepsy evaluation, prior knowledge about seizure onset zone localization is routinely available (from semiology, EEG, PET, etc. ). Leveraging this non-imaging information to guide a deep learning model is a logical and underexplored direction in FCD detection. This bridges the gap between purely data-driven methods and clinical workflow. 2) The adaptation of the GuideDecoder fusion mechanism from 2D vision to cortical surface meshes (using HexUnpool and SpiralConv) is a non-trivial technical contribution. It demonstrates a careful handling of irregular graph-structured data. The distance-based gating is a simple yet highly effective inference-time post-processing step. It cleverly re-purposes an existing auxiliary output (distance regression) to address the specific failure mode (increased false positives on healthy controls) of the text-conditioned model. The quantitative impact on specificity is substantial (from ~27% to ~93% on the main cohort). 3) The evaluation strictly follows the established MELD protocol, using a large multi-center dataset with a held-out independent test cohort (Bonn). This ensures a fair comparison and assesses generalizability. The use of multiple text conditioning variants (hemisphere only, lobe only, hemisphere+lobe) and the robustness strategy of random sampling during training (including “unclear” prompts) is a thoughtful experimental design that explores the granularity of required prior information.
- 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) The text prompts are derived from atlas-based overlap with the ground-truth lesion. In a real clinical scenario, the prior information would be provided as free text by a clinician (e. g. , “suspected left temporal lesion”). The current evaluation uses structured, clean atlas labels. The authors correctly note this as a future direction (free-form radiological narratives). It would strengthen the discussion to explicitly state that performance may degrade with noisy or imprecise clinical text and that this evaluation represents an upper-bound or “oracle” setting for text guidance. 2) The ablation in Table 4 focuses on gating aggregation rules, but does not isolate the effect of the text guidance itself versus the modified decoder architecture. The baseline is MELD. The Vision+LM model introduces both text fusion and the GuideDecoder with HexUnpool/SpiralConv. Could some of the gain be attributed solely to architectural changes? 3) The text mentions replacing standard 2D upsampling with HexUnpool and standard convolution with SpiralConv. A brief justification for why HexUnpool was chosen over simpler interpolation on the mesh would be helpful for readers unfamiliar with spherical surface processing. 4) In the References section, reference [8] (Park et al. ) appears to be duplicated or mis-cited. It is listed as a narrative review on AI for FCD, but the citation in text is for RadBERT. The actual RadBERT citation is missing. The authors should ensure the RadBERT model is properly cited. 5) The dice metirc should be used in the experiments to evaluate the performance of the proposed method.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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 paper presents a well-motivated, technically sound, and rigorously evaluated extension to the state-of-the-art MELD framework for FCD detection. The integration of clinically available prior knowledge via text guidance is both novel and practical. The proposed distance-based gating mechanism is an elegant solution to the specificity degradation introduced by the text conditioning. The manuscript is clearly written, and the experimental design adheres to high standards (multi-center data, independent test cohort, bootstrap confidence intervals). The identified weaknesses primarily concern presentation clarity, citation accuracy, and a more explicit discussion of limitations regarding text prompt realism.
- 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 explained the issues that I am conerned. I have no other questions.
Review #2
- Please describe the contribution of the paper
The authors propose a novel method for automated FCD detection in focal epilepsy from MRI, that combines the MELD graph neural network of cortical surface features with a text encoder of prior knowledge of FCD location. To deal with increased detection of false-positive clusters, the authors further introduce a gating mechanism of aggregated distance within each cluster to its boundary.
The model is trained and evaluated on participants of the MELD project. Performance of the model using different levels of coarseness in text input and gating mechanisms is compared to the standard MELD model in test and validation cohorts.
- 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 main strength of the paper is its potential in focal epilepsy to combine imaging-based input with other clinical information, such as radiology notes, which tend to be available for surgical candidates, but are currently not incorporated in automated lesion detection tools. In practice, pre-surgical evaluation considers a variety of sources of information, and the paper demonstrates that incorporating some of these into imaging-based localisation tools could substantially improve 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.
1.The authors motivate the incorporation of text-based information of lesion localisation with the typical availability of prior clinical knowledge. However, in their analysis the authors generate the text input by using the MELD pipeline to predict a lesion localisation, which is then mapped to an anatomical template to generate textual descriptions. This appears to be a circular way of generating input, which is then used alongside the MELD surface features to refine localisation. What is the conceptual justification for generating text input using a model based on MRI cortical surface features, only then to combine the text input with the same model of cortical features? Instead, why not use input independent of the MELD pipeline, e.g. notes of a radiologist, or findings of EEG? 2.Another methodological consideration is the gating mechanism. Does applying a median (min/max) threshold of the distance of vertices within a cluster to its boundary not bias towards detecting small and, in particular, narrow lesions? What is the justification for using this mechanisms specifically? 3.Further, table 4 compares the novel method with all gating mechanisms to MELD with the same gating options. However, the discussion of this is very limited, e.g. the authors do not discuss that MELD with added gating mechanism outperforms their model in three out of five metrics. It is also unclear why these results are only show for the main cohort, and not the independent validation cohort.
- 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.
- 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?
The paper makes a novel and interesting contribution. Whilst the choice of some analysis steps are unclear, the results are convincing, and the method has clinical relevance with potential further development in the future.
- 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.
My concern regarding the apparent circularity of using MELD-derived predictions as text input has been clarified, in that the text descriptions are derived from ground-truth lesion masks, not model output. The authors also address the gating mechanism concern convincingly, noting true-positive clusters are preserved regardless of lesion size. The point about MELD+gating outperforming Vision+LM on some specificity metrics is reasonably framed as a sensitivity-specificity trade-off.
Review #3
- Please describe the contribution of the paper
The work proposes a multimodal segmentation framework that integrates surface-based visual features with text descriptions that encode prior knowledge on anatomical localization of the seizure onset to guide Focal Cortical Dysplasia (FCD) detection.
- 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 work tries to combine surface-based and textual data to segment lesions in the brain
- 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 work is not clear organized. It is hard to follow and the motivation is not clear. What advantage does the text adds to the segmentation?
- The authors call it segmentation but it should rather be parcellation as it is surface-based.
- These text-based inputs are very primitive and does not provide enough information as the results are quite inferior in general; huge portions of the legions are missed.
- Why was surfaces used? Can we use volumetric segmentations (without the graph-based framework) and then check the results with that; perhaps an ablation study might be good?
- The gating mechanism lacks technical grounding what is the use of computing the SDF?
- The different loss functions need ablation studies to see what is the true effect of each loss.
- The results are poor as the overlap is very small (red vs black outline)
- The work needs to be evaluated on different domains.
- You only show segmentation in one limited region
- Please rate the clarity and organization of this paper
Poor
- 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.
(1) Strong Reject — must be rejected due to major flaws
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- Limited evaluation on only one dataset in a very localized region and lack of clear technical description
- It is not clear merely integrating two modalities can produce any improvements in the segmentation
- Reviewer confidence
Very confident (4)
- [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 rebuttal still doesn’t address the main issues. The evaluation is very limited and the concern regarding the simplicity or the lack of adequate text prompts have not been addressed. The concern regarding the overlap of the segmentation is not addressed: the red and black overlap between the prediction and the groundtruth is way too off.
Author Feedback
We thank the reviewers and meta-reviewer. The meta-review highlights three themes (text generation, gating, evaluation choices), grouped below.
[1] Text input: source and realism (R1, R2, MR) R2 raises a concern that the text appears circular (MELD predictions re-fed into MELD). We respectfully clarify this is not the case: textual descriptions are derived from ground-truth lesion masks registered to template space via AtlasReader (Sec. 3.1), following MELD’s localization procedure but using GT, not model output. The goal is a methodological proof of concept that text representations of coarse anatomical priors can guide detection. We agree this setup is an upper-bound evaluation, performance may degrade with noisy clinical text, and radiology notes or EEG findings are the natural next step. Such data was not consistently available across the 23-center MELD cohort. This is discussed in Sec. 5 and we will make it more explicit, as requested by R1. [2] Distance gating: rationale and bias (R2, R3, MR) The gating uses the existing MELD signed-distance head (Sec. 2.3): inside lesions distances are negative, outside they are positive and grow with distance from the boundary. FP clusters in healthy cortex retain high positive distances and are filtered, while true-lesion clusters contain near-/negative-distance vertices and are kept. The Youden’s J threshold is positive on our data; therefore, as long as distance predictions are accurate, true-positive clusters are not filtered regardless of lesion size. Empirically, sensitivity is largely preserved on both cohorts (Tables 2-3) while specificity rises from 27.5% to 93.3% (main) and 37.3% to 94.0% (Bonn).
[3] Evaluation choices and ablations (R1, R2, R3, MR) R3 states evaluation is on a single dataset. We respectfully note Sec. 3.1 describes two cohorts: the multi-center MELD cohort (23 centers, 1.5T/3T) and the independent Bonn cohort as external test set. Improvements over MELD are consistent on both, supporting generalizability. R2 notes MELD+gating outperforms Vision+LM on some specificity metrics (Table 4). Our objective is improved sensitivity, the clinical bottleneck in FCD detection. Vision+LM detects more lesions and yields higher IoU, particularly on Bonn (sensitivity 78.0% vs 69.5%, IoU 0.306 vs 0.218); we view this as a sensitivity-specificity trade-off and will frame it as such in the revision. R1 asks whether gains stem from text or from the modified decoder. In an unreported ablation, Vision+LM with “unclear” prompts at inference (no localization content) shows no consistent gains over MELD; gains appear only with informative prompts, indicating the contribution is not purely architectural. If permitted, we will add a note on this ablation to the final version.
Reviewer-specific points R1: HexUnpool preserves the icosphere mesh correspondence during upsampling, ensuring consistent feature propagation across resolutions and integrating naturally with SpiralConv, unlike generic interpolation. The RadBERT citation was a preparation error and will be corrected. IoU and Dice are monotonically related (Dice = 2*IoU/(1+IoU)); we used IoU per the MELD protocol for direct comparability.
R3: The surface formulation follows MELD and is motivated by FCD manifesting as subtle cortical morphological/signal changes along the cortical sheet; the model performs vertex-wise lesion prediction, a segmentation task in surface-based neuroimaging, not parcellation. The loss exactly follows MELD (Sec. 3.2) for fair comparison. IoU values are in line with published MELD on this challenging task; the relevant signal is consistent improvement on both cohorts. We will improve methods clarity and will release code upon 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.
The paper proposes the ingenious idea to combine text and images for the detection of FCD. The experimental design is performed on a multi-centre dataset with efforts to assess generalisability. There are certain aspects that would benefit from better clarification / justification in a rebuttal especially with respect to the text input and its generation as well as the gating mechanism. Choices for the evaluation in terms of independent cohort and ablation investigations should also be discussed and justified if possible.
- 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 authors have well addressed all comments and clarified important points yielding a much stronger submission that should have its place in the conference.
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.
After rebuttal, R1 and R2 remained Accept, while R3 maintained Reject, due to that the evaluation remained limited, the text prompts were overly simplistic and unrealistic, and the proposed method offered limited methodological novelty beyond integrating text guidance into the existing MELD framework. While incorporating clinical prior knowledge is an interesting direction, the current evidence is insufficient to demonstrate a substantial technical advance or broad generalizability, leading to my recommendation of Reject.
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 has been a difficult choice as I have some mixed feelings after reading the paper and checking all comments post-rebuttal. On the positive side, I think the main inspiration for the proposal is sound and interesting. If for pre-operative scans there is prior knowledge from a radiology report, why not use it? On the negative side (and this is mostly highlighted by reviewer #3 in their original comments and post-rebuttal), the evaluation has some issues and the final implementation of “texts prompts” is questionable. I am not entirely convinced by the rebuttal, especially on the “text prompt” part and that may be caused by the fact that I have seen other papers with LLMs that suffer from similar issues. The advantage of model for radiological reports is their ability to extract meaningful information from unstructured text reports. The “text prompt” here are simple and essentially amount to categorical variables (could just be one-hot encoded vectors). The use of a LLM for the current implementation is not well justified.
However, and this is where my final decision comes from, the authors agreed to contextualise their work and explain that the results might not be representative to a real setup with text reports and for this paper in particular I think the strengths and novel-ish idea is good enough for acceptance. I would still advise the authors to extend their current work with real radiological reports for a more impactful work.
