acceptodds
Under review as a conference paper at ICLR 2027

BeViL: Brain Tumor Detection Beyond Visible Lesions Using Secondary Effects via Contrastive Loss

Abstract

Accurate brain tumor detection from Magnetic Resonance Imaging (MRI) remains challenging when lesions are subtle, partially visible, or missed due to imaging artifacts, sparse slice acquisition, and reconstruction errors. Most existing methods primarily rely on visible lesion appearance and feature fusion across modalities, which require high-resolution, artifact-free, and multimodal 3D MRI acquisitions, which are often time-consuming and costly. To address this limitation, we introduce a novel problem formulation for brain tumor detection that relies solely on the tumor-induced secondary effects present in neighboring slices. Using paired FLAIR, T1-weighted, and T2-weighted MRI, modality-specific DenseNet201 classifiers are trained with a joint objective that combines cross-entropy loss and supervised contrastive loss. The contrastive objective aligns class-consistent representations across modalities, enabling effective representation learning by encouraging shared discriminative features that capture indirect tumor-related cues beyond explicit lesion visibility. During inference, each modality is employed independently. Experimental results demonstrate that the proposed framework consistently outperforms single-modality baselines and non-contrastive multimodal training strategies. Using only secondary tumor-induced effects rather than explicit lesion appearance, the proposed approach reduces the risk of misdiagnosis in the case of sparse acquisitions and can be extended to more reliable and early detection of brain tumors.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multimodal Learning, Brain MRI, Secondary Effects, Lesion Detection, Contrastive Loss

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_17mfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper talks about a problem of identifying visually non-evident tumors by leveraging the secondary structural effects on adjacent slices. This scenario is particularly relevant in clinical settings with low-resolution or thick-slice CT/MRI scans, where subtle lesions are frequently missed. To simulate sparse sampling in real-world practice, the authors construct a dataset from public 3D brain MRI scans, designating non-visible slices within 1–3 slices of a lesion as positive, and those 4, 6, or 8 slices away as negative. The experimental evaluations demonstrate that the proposed approach effectively reduces the rate of missed diagnoses.

Critical strengths

A positive aspect of this work is its effort to formulate a new problem, specifically detecting non-visible tumors via secondary structural manifestations in adjacent slices. This attempts to model real-world clinical scenarios involving sparse sampling and thick-slice scans. While I have concerns regarding other parts of the paper, the authors' initiative to observe such a clinical scenario is appreciated.

Critical weaknesses

The authors explore a novel and clinically meaningful question, which is commendable. However, the link between the experimental setting and the intended clinical scenario requires further clarification. Tumors most likely to be missed because of slice thickness, interslice gaps, or partial volume effects are generally those with a limited extent along the axial direction. It remains unclear whether such small lesions produce sufficiently consistent and detectable secondary effects, such as edema, mass effect, or tissue deformation, on adjacent slices that do not directly contain the lesion. Conversely, larger tumors with prominent secondary effects typically span multiple slices and may therefore be less likely to be completely missed under sparse acquisition. This is the main concern regarding the research question of the paper.

Questions for authors
  1. Although the research question is interesting, the technical novelty of the proposed approach is not convincingly established. Most of its components are standard and have been widely adopted in existing studies. The authors should clearly identify the methodological contribution beyond the combination of established techniques and substantiate it through rigorous comparisons and ablation studies.

  2. The current description of the methodology is insufficiently detailed. The manuscript does not provide a clear and formal account of the complete workflow from data input to the predictions generated by the four classification heads. Based on Figure 2 and the accompanying text alone, it is difficult to determine how features are extracted and propagated, how the classification heads are defined and supervised, and how their outputs are combined. The authors should substantially revise this section by providing explicit mathematical formulations, notation, feature dimensions, and complete details of the training and inference procedures.

  3. Figure 3(b) is not sufficiently clear to allow a meaningful interpretation of the results. It should be redrawn with improved resolution, clearly labeled axes and legends, distinguishable colors or line styles, and a more readable overall layout.

  4. The English presentation requires some improvement. The authors may want to substantially revise the article to allow the technical content to be communicated accurately.

Reviewer_NMPmOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The authors propose BEVIL to address the risk of missing small brain tumor lesions in MRI. The method introduces a new problem formulation for brain tumor detection that relies solely on tumor-induced secondary effects in neighboring slices and is optimized using cross-entropy and supervised contrastive losses. Experimental results demonstrate promising performance, suggesting potential for more reliable and earlier detection of brain tumors.

Critical strengths
  1. BEVIL infers tumor presence from tumor-induced secondary effects in neighboring slices, without requiring the lesion to be directly visible in the target slice.
  2. The paper explores a new problem of tumor detection based solely on secondary effects, which has potential research value.
  3. The method could provide tumor-aware priors for reconstructing 3D MRI volumes from sparsely sampled 2D slices.
Critical weaknesses
  1. In the Experimental Setup, slices within three slices of the tumor are labeled positive, while selected slices farther away are labeled negative. How can the authors ensure that these negative slices contain no secondary effects? Conversely, neighboring slices may not exhibit identifiable secondary abnormalities. Therefore, the current experiments directly evaluate the ability to distinguish nearby from more distant slices according to a predefined distance rule, but do not establish the claimed ability to identify secondary effects or detect tumors at the patient level.

  2. The architectural novelty is limited, as the framework combines existing components, including DenseNet201, an MLP projection head, and a standard supervised contrastive objective. Moreover, the positive sets in Equations (3)–(5) are defined solely by class labels, consistent with existing SupCon formulations. More importantly, the distinct contribution of using paired views of the same underlying anatomy remains unclear.

  3. The external-test ROC-AUC values in Figure 3 appear inconsistent with the sensitivity and specificity reported in Table 1. Were these external evaluation metrics computed from the same set of predictions?

  4. The model is fine-tuned for five epochs on 10% of the patients in UTSW-GLIOMA and then evaluated on the remaining patients. This constitutes evaluation after target-domain adaptation. Although this protocol is common, it does not directly demonstrate generalization to an external institution without adaptation. The authors should report results both with and without fine-tuning and clarify whether the comparison methods were given the same adaptation opportunities.

  5. The dataset description and experimental details are insufficiently clear. In particular, the manuscript does not adequately specify the validation split within the development set, the criteria for selecting hyperparameters and the slice-gap setting, or the pretraining and fine-tuning procedures used for the main baselines.

Questions for authors

See Weaknesses.

Reviewer_pXpsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces BeViL, which aims to classify MRI slices using indirect imaging cues in slices neighboring tumors. The framework uses a dedicated DenseNet encoder and classification head for each of four modalities, together with a shared projection head, and is trained jointly with classification and supervised contrastive losses. The experiments construct near- and far-from-tumor slice labels using tumor annotations from complete volumes. Evaluation is conducted on BraTS 2023, with additional external testing on UTSW-GLIOMA following fine-tuning on a subset of that cohort, as well as a small-scale evaluation on healthy T1-weighted scans.

Critical strengths

The multimodal-training, single-modality-inference design is appealing. Retaining a separate classification branch for each modality allows the model to benefit from complementary information during training without requiring all four modalities at inference. This offers greater flexibility than conventional multimodal fusion approaches that require multiple modalities at test time.

Critical weaknesses
  1. The evaluated task does not adequately address the problem posed by the paper. The stated motivation is to detect tumors through indirect signs when MRI acquisition does not capture the tumor itself. However, the experiments primarily distinguish slices near tumors from those farther away. These are not equivalent tasks. Evidence is still needed to show that, when the acquired images omit the tumor itself, the remaining secondary effects are sufficient for reliable patient-level tumor detection. I therefore remain unconvinced that the current experiments adequately support this motivation.

  2. The method’s independent contribution remains unclear. Although the paper reports improvements over classification baselines, it should clarify whether the comparisons use matched training conditions. The modality ablations also change the number of training modalities and therefore do not isolate the contribution of supervised contrastive learning. Furthermore, positive pairs are defined solely by class labels, and the additional benefit of exact multimodal correspondence at the patient and slice levels has not been evaluated separately.

Questions for authors
  1. Can the current slice classification task serve as a proxy for detecting tumors not captured during acquisition? Distinguishing slices near tumors from those farther away is not equivalent to detecting a tumor that is absent from the acquired MRI slices. Please specify the intended acquisition scenario and explain how the experiments represent that scenario. If direct validation is unavailable, the claims should be limited to the evaluated slice classification task.

  2. What are the respective benefits of supervised contrastive learning and exact multimodal pairing? Table 2 includes classification baselines, but whether their training conditions are matched needs clarification. Defining positives solely by class labels also does not establish the value of exact pairing. Please clarify whether the existing baselines provide matched no-contrastive-loss controls and provide evidence for any additional benefit of pairing modalities from the same patient at the same slice position.

  3. Does T1CE participate in supervised contrastive learning? The architecture diagram and total loss involve four modalities, whereas the contrastive-learning description refers to only three feature extractors and uses (N=3B). Please clarify whether T1CE contributes to the contrastive loss, specify the actual batch composition, and ensure that the modality count and definition of (N) are consistent across the text, equations, and figure.

  4. What evidence supports the claim that the model relies on secondary effects? Grad-CAM provides useful clues but does not rule out reliance on anatomical location or other correlated features. Please provide more direct evidence, for example, by comparing model responses with independent expert annotations of edema, mass effect, or related signs. Without further validation, this mechanism should be presented as a possible explanation rather than an established finding.

Reviewer_fvfyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BeViL investigates whether tumor presence can be inferred from secondary effects in neighboring MRI slices that do not directly contain the tumor. The method jointly trains four modality-specific DenseNet201 branches using classification and supervised contrastive losses. Each modality can be used independently at inference. The authors derive distance-based labels from BraTS2023, assigning positive labels to slices 1–3 slices away from the tumor and negative labels to slices at distances 4, 6, and 8. Evaluation includes an internal patient-level test split, external testing after fine-tuning, and an additional healthy-subject evaluation.

Critical strengths

1.The problem is meaningful. Extracting information from slices without direct tumor visibility could benefit sparse-slice analysis and pathology-aware reconstruction.

2.Inference requires only one modality. Paired multimodal information is used during training, while individual branches operate independently at deployment.

3.The reported classification improvements are substantial. In Table 2, FLAIR accuracy increases from 84.69% with the single-modality DenseNet201 baseline to 91.56%. Improvements are also reported for other modalities. The manuscript specifies a patient-level internal split and includes modality ablations and different distance settings.

Critical weaknesses

1.The distance-based labels do not establish secondary-effect detection. Positive and negative slices both come from tumor patients and differ by distance from the annotated tumor. Secondary effects are not independently labeled, and the segmentation regions used to define tumor-containing slices are unclear. Anatomical position or other distance-related cues could explain the results, while Grad-CAM does not rule out these alternatives. Please clarify lesion-exclusion criteria and provide blinded image assessment and anatomically matched controls. See the labeling protocol on page 5 and Figure 5.

2.The evaluation does not establish sparse-acquisition or patient-level detection performance. Slices are selected from complete volumes without directly testing acquisition spacing, sampling offsets, or artifacts. External results follow fine-tuning on 10% of UTSW patients, and the healthy-subject evaluation includes only 20 T1W scans. Please report patient-level performance under simulated sparse acquisition and external results before and after fine-tuning. Confidence intervals should also account for correlated slices within patients. See Sections 3, 5, and 6 and Tables 1 and 4.

3.The training description is inconsistent, and the ablations do not isolate contrastive learning. The architecture has four branches, but Section 4 refers to three feature extractors and defines (N=3\times\text{batch size}), listing only T1, T2, and FLAIR. Equation (3) defines positives by class, leaving the role of paired correspondence unclear. Please specify whether T1CE participates in contrastive learning and add a four-branch classification-only control under identical training settings. The validation split and hyperparameter selection procedure also need clarification.

Questions for authors

1.Which segmentation labels define the tumor-containing slice set (T)? Does it include edema? Were positive slices independently reviewed to confirm the absence of directly visible lesions? 2.Does performance persist after matching axial position and tumor location? What evidence excludes reliance on anatomical-position cues? 3.What is external performance without fine-tuning? What patient-level sensitivity and false-positive rate are obtained under simulated sparse acquisition? 4.Does T1CE enter the supervised contrastive loss? How does the four-branch model perform using classification losses alone?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.