acceptodds
Under review as a conference paper at ICLR 2027

MUGAT-KD:A Multimodal-to-Unimodal Knowledge Distillation Framework for Pulmonary Nodule Malignancy Classification

Abstract

Pulmonary nodule malignancy classification is a vital component of early lung cancer screening. Multimodal models that integrate CT images with structured nodule attributes can derive more comprehensive nodule representations; however, the inference process relies on additional attribute annotations, limiting their practical applicability. Existing distillation methods primarily align model outputs or global representations, making it difficult for a unimodal model to preserve the hierarchical influence of multimodal knowledge on feature representations, attention to critical slices, and local regions of interest. To address this limitation, we propose MUGAT-KD, a multimodal-to-unimodal knowledge distillation framework. The framework employs a graph attention network to model the spatial relationships among CT slices and the associations between slices and nodule attributes. A frozen multimodal teacher then guides the student model at three levels: global features, slice-level attention, and saliency-aware patch masking. We conduct within-dataset classification experiments on LIDC-IDRIandLIDP datasetsandevaluate cross-dataset generalization by training on LIDC-IDRI and performing external validation on LIDP. Theexperimentalresults demonstrate that MUGAT-KD effectively transfers attribute knowledge from the multimodal teacher to the unimodal student, improving pulmonary nodule malignancy classification without requiring additional nodule attribute annotations and enhancing adaptability to differences in data standards. Ablation studies verify the complementary effects of the distillation modules. Attention visualizations further show that the student model learns the multimodal teacher’s attention patterns over critical CT slices and local regions. These results establish MUGAT-KD as an effective solution for pulmonary nodule malignancy classification when nodule attributes are unavailable.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Analysis, Knowledge Distillation, Multimodal Learning, Graph Neural Networks, Pulmonary Nodule Classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YCqhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a new method that classifies pulmonary nodules malignancy on a binary basis (malignant or benign). For training the authors use input that is labeled both on malignancy and that contains several attribution labels. These attribution labels are generally considered helpful for classification. In the training setting a teacher network is trained using these helpful attributes and the CT scans and a student network is trained to classify only based on the CT scans. Information from the teacher network is distilled to the student network based using several methods such as a distillation loss. The contribution is a new method for CT pulmonary nodule classification using multimodal-to-unimodal hierarchical knowledge distillation and graph neural networks. Validation is done using two different datasets. Both datasets are evaluated using five-fold cross-validation and external validation is being done by training the network on one dataset and evaluating it on the other one.

Critical strengths

Valuable addition to the landscape: The paper is novel in its application rather than its methodology: it brings knowledge distillation to a domain that was previously dominated by CLIP-style models. Thus, it fills a valuable gap in this domain.

Critical weaknesses
  1. No clear methodological contribution: For a venue like ICLR I would expect a clear methodological contribution. This contribution in itself is quite sound, but the contribution is rather by using existing methods on a new domain than by providing a new method of tackling a problem. All methods of this contribution are sound, but none seems to be new. Sometimes this is stated in the paper and sometimes it is not. I would also expect a contribution that goes beyond one specific domain and application case. I would suggest to rather go for a medical venue like e.g. MICCAI. For a methodological contribution I would rather expect a comparison to knowledge-distillation methods in the same domain, to see whether this methods outperforms other knowledge-distillation methods.

  2. Some results are non-comparable: a) Some of the results were copied from other papers and not replicated. This, by itself is not an issue, but the other papers partly used different evaluation and exclusion criteria e.g. results from Nodule-Clip were copied. However, Nodule-Clip had a more difficult learning and evaluation setting, since they did not exclude any nodules (they classified all scores below and equal to 2.5 as benign and above 2.5 as malignant). This paper has a different setting and excludes the indeterminate nodules (mean score exactly 3), which are likely among the hardest to classify. This makes the comparison between these methods incomparable. I just checked this for Nodule-Clip, but this might as well be the case for other methods (notably it is not for CLIP-Lung, which seems to be replicated). I understand, that it is not always possible to replicate all the findings and it is even stated, that some values were copied (not stated which ones though). But this comparison makes it seem as if MUGAT-KD outperforms Nodule-CLIP on every metric, while the honest comparison would be, that they are non-comparable. An ablation using different thresholds and including all datapoints might solve this issue and provide some fair comparison.

b) This is a weaker issue than above, but also comparing MUGAT-KD with 2D and 3D CNN based method is without further comment not sound. This comparison is interesting, however these methods were trained without any attributes and thus with a different training setting. I see the usefulness of them as a type of "lower bound", but the setting is fundamentally different. I agree with putting them there, but that should also be explicitly stated. In the same breath, the teacher (Tab. 3) uses the attributes at inference and thus acts as an "upper bound", but it is not compared to any other method that uses the attributes at inference (e.g., the ones discussed in Sec. 2). It would be interesting to see how the teacher performs against them.

To conclude: I think for now the method is truly only comparable with CLIP-Lung among the listed comparisons. This is however nowhere explicitly stated.

Questions for authors

There are no comparable methods for the cross-dataset evaluation method. Since CLIP-Lung is already implemented, could you do a comparison at least with this method to see how it performance compares. I would also be interested in how inputs are encoded. This is at least for the attributes never listed. How were the teacher checkpoint and the epoch thresholds of Eq. 16 chosen, given that the five folds are reported as test folds? And how many nodules (benign/malignant) remain in LIDC-IDRI after excluding score-3 nodules?

Here just some minor issues (no effect on score), which you could improve: The paper has several spelling errors and some formatting errors(these are just the ones I found maybe use a LLM to parse for all of them) in Fig. 2(b), "Maskd CT Slice" → "Masked CT Slice" and "Masked Central Node Neature" → "Feature"; in Fig. 4, "Probabilitles" → "Probabilities"; in Sec. 4.1, "benign andmalignant,respectively" is missing spaces, and "Exclusion with a score equals 3" is ungrammatical (e.g., "Nodules with an average score of exactly 3 are excluded"); in the Conclusion (l. 480–481), the line break after "feature-" leaves the comma at the start of the next line (", slice-, and region-level").

1 comment

TO Critical Weakness 1 : Our contribution is a generalizable multimodal-to-unimodal distillation framework for classification tasks involving 3D images and structured data, rather than a method limited to medical imaging. We represent a 3D image as a sequence of slices and construct a multimodal teacher with a slice-attribute graph that models relationships between image slices and structured attributes. An image-only student then learns through three complementary objectives: global representation alignment, slice-level attention alignment, and saliency-guided masked distillation. Unlike CLIP-Lung, which mainly aligns global image and attribute representations, our framework transfers knowledge at the representation, slice-attention, and region-conditioned levels. It assumes that structured attributes have observable manifestations in the images and is therefore potentially applicable to other 3D imaging tasks with complementary structured or multimodal information.To establish a new paradigm for applications in computer vision, audio, language, and other modalities. TO Critical Weakness 2.a: We have now marked results reported in the literature with asterisks (*). Table 13 in the supplementary material summarizes the dataset sizes, class distributions, and label definitions of all methods marked with an asterisk. The remaining methods in Table 1 and all methods in Table 2 were reproduced using the same patient-level data split. We have also clarified that LIDC-IDRI labels are based on radiologists’ malignancy ratings, whereas LIDP uses pathology-confirmed diagnostic labels. TO Critical Weakness 2.b: The training information available to MUGAT-KD differs from that available to conventional 2D/3D CNN methods. These CNN methods use only CT images during both training and inference, whereas MUGAT-KD leverages structured attributes as privileged information during training while requiring only CT images at inference. We retain these methods to provide reference performance under image-only training conditions and to assess the benefits of incorporating attribute knowledge during training. As the reviewer noted, these methods serve as a performance lower bound. We will explicitly clarify the scope and interpretation of these comparisons in the revised manuscript. Furthermore, in Table 12, we provide a comparison between our teacher model and other multimodal methods, serving as a performance upper-bound reference. TO Question:

  1. We have now reproduced the LIDC-to-LIDP cross-dataset evaluation using the publicly available CLIP-Lung implementation, with the results reported in Table 10 of the supplementary material. Furthermore, since the experiment in Table 4 involved training and validating a single model on the full LIDC dataset before evaluating it on LIDP, the reported results may be subject to uncertainty. To address this, we additionally conducted five-fold cross-validation on LIDC, training five models and evaluating each independently on the entire LIDP dataset. We report the mean performance and sample standard deviation across the five evaluations in Supplementary Table 10. 2.The structured input consists of nine attributes: internal structure, calcification, subtlety, sphericity, margin, lobulation, spiculation, texture, and nodule length. Each attribute is first converted into a discrete integer index and then mapped to a 64-dimensional vector through an independent learnable embedding layer. These vectors serve as attribute nodes and are jointly fed into the graph attention module along with the CT slice features. 3.We do not use the five folds as separate test folds. A total of 1,454 nodule samples were included. We first established a fixed test set of 242 samples using patient-level splitting. The remaining 1,212 samples were divided into five validation folds, each containing 242 or 243 samples. In each run, four folds were used for training (969 or 970 samples), one fold for validation, and evaluation was always performed on the same fixed test set. The best teacher checkpoint was selected based on the validation AUC within the first 75 epochs. Epochs 75, 100, and 150 were predefined as the starting points for teacher freezing, attention distillation, and masked distillation, respectively. The best student checkpoint was selected based on validation performance after epoch 75. These settings were consistent across all folds, and the test set was never used for checkpoint selection.

We thank the reviewer for pointing out these spelling and formatting issues. We have carefully checked and corrected each of the identified errors and conducted a further review of the language and formatting throughout the manuscript to improve its accuracy and readability.

Reviewer_UmUPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The method trains a multimodal graph teacher with CT slices and structured nodule attributes, then trains a CT-only graph student using global-feature, slice-attention, and saliency-aware masking supervision. Tables 1–2 report AUCs of 98.15 on LIDC-IDRI and 81.00 on LIDP. External LIDC-to-LIDP validation reaches 75.28 versus 68.84 for image-only training, a 6.44-point improvement. LIDC labels threshold radiologist malignancy scores and are not uniformly pathology-confirmed diagnoses; LIDP uses pathological outcomes.

Critical strengths
  1. The application setting is practical and includes external validation. Missing attributes at inference are a relevant constraint. The work evaluates across datasets rather than reporting only an internal LIDC split.
  2. The distillation components are explicit, and code is supplied. Table 3 examines component combinations, and Table 4 includes cross-domain combinations. The paper explicitly states patient-level five-fold splitting, so nodule-level leakage should not be alleged without evidence.
Critical weaknesses

W1: The released evaluation does not reliably implement the central CT-only student evaluation, and configurations materially disagree with the paper

The following are statically verifiable artifact problems. They do not establish that the reported paper results were necessarily generated by this erroneous path.

  • In evaluate.py:426–436, the default branch returns pred, mid3, loss, but predictions and AUC use argmax(mid3) and softmax(mid3). In models/mugat_kd.py:594–596, mid3 is assigned teacher logits; the student output is out. The default configuration has only_img=False, and evaluation sets epoch=30, entering the attribute-assisted teacher path rather than scoring the CT-only student.
  • With only_img=True, the encoder assigns mid3=None around models/mugat_kd.py:558–561, while evaluation still applies argmax to it. The teacher-only and return_student_attention branches also do not consistently assign the mid3 later used for scoring. Simply enabling only_img does not fully repair this script.
  • Equation 16 starts mask distillation at epoch 150; config.py:14 sets 175. For epochs 75–100, the manuscript uses 0.9 classification + 0.1 feature loss, whereas the implementation around line 715 uses 0.8 + 0.2.
  • The manuscript, p. 8, gives masking probabilities 0.15/0.35/0.65; models/mugat_kd.py:660–669 uses 0.1/0.3/0.6.
  • The manuscript describes a four-block slice CNN. main.py defaults to a residual encoder, although --image_encoder slice_cnn is supported. Without exact reproduction commands, the encoder used for the reported tables is unclear.
  • The ZIP does not include patient-grouping/split JSON files, prediction logs, or checkpoints. train.py reads precomputed fold files; static inspection cannot verify the patient-level guarantee. This absence alone does not demonstrate leakage.

These issues concern the central claim that the student achieves the reported results without attributes, rather than peripheral code quality. The authors should provide the exact version, commands, and per-case student predictions for Tables 1–4 and demonstrate unchanged inference metrics after attribute inputs are removed entirely.

W2: Controls do not sufficiently isolate the three-level mechanism or establish comparison fairness

  • Table 3 shows gains from 93.73 to 98.15 AUC on LIDC and 74.31 to 81.00 on LIDP. However, matched-budget controls for ordinary KD from a CT-only teacher, logit-only KD, and random/uniform masking are absent. Feature/slice/mask combinations alone do not separate hierarchical multimodal knowledge from generic teacher supervision, augmentation, or training duration.
  • Page 8 mixes reproduced public methods with published numbers without marking Table 1 row by row. Rankings across different patient splits and evaluation protocols cannot provide strong SOTA evidence.
  • Separate validation within five-fold evaluation, teacher checkpoint selection, final included nodule counts and class distributions, spatial resampling, and HU processing need clarification. Pre-cropped .npy inputs and hard-coded paths do not supply those missing records.
  • LIDP specificity is 44.73 versus CLIP-Lung's 58.53. Higher sensitivity is useful, but does not by itself establish a superior clinical trade-off. Validation-selected thresholds, specificity at fixed sensitivity, or ROC/PR analysis would be appropriate; a new large clinical trial is not requested.
Questions for authors

see weaknessa

1 comment

We sincerely thank the reviewer for carefully examining our code and attempting to reproduce our experiments. We acknowledge that the identified discrepancies are valid. After completing our main experiments, we continued exploring additional hyperparameter configurations but inadvertently failed to restore the original settings before submission. We apologize for this oversight. We have now corrected the code and uploaded supporting evidence. Our point-by-point responses follow.

Response to Critical Weakness 1: Code and Reproducibility (1) The teacher_only flag is used to train the teacher independently, whose performance is reported as No. 0 in Table 3. During joint teacher–student training (teacher_only=False), we temporarily used argmax(mid3) to inspect the pretrained teacher's performance. We have corrected the evaluation code and added a comment explaining that pre should be replaced with mid3 only when evaluating the teacher. (2) only_img Evaluation. This issue arose from the same modification. We changed argmax(mid3) and softmax(mid3) in evaluate.py to evaluate the teacher. The default configuration was correct, but when only_img=True, mid3=None, making these operations invalid. This issue has now been corrected. (3) Training Schedule and Loss Weights. We tested starting masked distillation at epoch 175 because we were concerned that slice-level attention distillation might not have converged by epoch 150. The results showed convergence before epoch 175. The loss-weight discrepancies similarly resulted from testing three configurations (feature / attention / mask):

C1:(0.9, 0.1, 0, 0) / (0.8, 0.1, 0.1, 0) / (0.7, 0.1, 0.1, 0.1) C2:(0.8, 0.2, 0, 0) / (0.8, 0.1, 0.1, 0) / (0.8, 0.1, 0.1, 0.1) C3:(0.7, 0.3, 0, 0) / (0.7, 0.2, 0.1, 0) / (0.7, 0.1, 0.1, 0.1) The corresponding results are provided in Supplementary Table 6. (4) We evaluated three settings: 0.15/0.35/0.65, 0.10/0.30/0.60, and 0.20/0.40/0.70. Results are reported in Supplementary Table 7.

(5)We trained and evaluated both image encoders, with results presented in **Supplementary Table 5.

(6) The updated code release now includes data split files, model checkpoints, and other supporting files to facilitate reproducibility and verification.

Response to Critical Weakness 2: Controls and Comparison Fairness (1) Additional Controls.We added the requested controls using identical student architectures, data splits, training budgets, and evaluation protocols. On LIDC-IDRI, the AUCs were 95.44% for CT-only teacher logit KD, 96.10% for attention-only KD, 96.63% for random masking, 96.90% for multimodal teacher logit KD, and 97.31% for simple attribute concatenation, compared with 98.15% for full MUGAT-KD.

These results suggest that the improvements cannot be attributed solely to generic distillation, stronger supervision, or masking regularization. Instead, feature-, attention-, and region-conditioned knowledge transfer provide complementary benefits. Results are reported in Table 3, Supplementary Tables 7 and 11, with further discussion in Section 4.3.

(2) Comparison Fairness. We revised Tables 1 and 2 to distinguish reproduced results from published values. Methods marked with an asterisk (*) were not reproduced because their source code was unavailable, whereas all unmarked methods were reproduced under identical experimental settings. We additionally document the dataset configurations of non-reproduced methods in **Supplementary Table 13 to improve transparency.

(3) Data Splits, Checkpoints, and Preprocessing. As documented in split1.json, we use an independent, fixed test set. For each fold, four development folds are used for training and the remaining fold for validation. The test set remains unchanged across all five folds. During the first 75 epochs, the best checkpoint is selected based on the teacher's validation AUC. The current LIDC splits contain 1,454 unique nodules (1,020 benign and 434 malignant). No spatial resampling was performed; the original voxel spacing was preserved. CT HU values were clipped to [-1350, 150], linearly normalized to [0, 1], and saved as .npy files. This preprocessing was not repeated during training.

(4) Clinical Trade-off on LIDP. We agree that higher sensitivity at the default 0.5 threshold does not establish clinical superiority, and we have removed this claim. In Supplementary Table 9, thresholds are selected exclusively on the validation set using Youden's index. The revised manuscript also explicitly discusses the sensitivity–specificity trade-off.

Reviewer_xiZROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MUGAT-KD transfers a CT/attribute graph teacher into a CT-only student through feature, slice-attention, and saliency-guided masking losses. The application is reasonable, but the contribution appears incremental, and the comparisons do not sufficiently isolate the value of the proposed hierarchy.

Critical strengths

The paper includes all component combinations and pathology-defined external validation. LIDC-to-LIDP AUC improves from 68.84% to 75.28% over the image-only model. These results support effectiveness within the evaluated setting.

Critical weaknesses
  1. The research setting and building blocks are established. CLIP-Lung [1] already uses attribute annotations during training without additional inference overhead. Attention Transfer [2], AttMask [3], and masked feature distillation [4] provide relevant precedents, although their implementations differ. MIE-GAT, cited by the authors, also motivates a clearer boundary for the graph contribution. The plausible novelty is the particular attribute-informed graph and distillation combination. Moreover, Eq. 15 aligns global center-node features after spatial masking; it does not explicitly match local features. Clarify what makes this region-level knowledge transfer beyond guided augmentation.

  2. The controls do not isolate the proposed mechanism. Table 3 includes feature-only distillation and all module combinations, which is valuable. However, under identical teachers, students, and training budgets, compare conventional output/attention distillation, random masks with matched masking rates, and a simpler attribute-concatenation teacher. An image-only teacher would further separate the benefit of attribute knowledge from generic distillation. Without these controls, gains could reflect stronger supervision or regularization rather than a distinctive hierarchical mechanism.

  3. The full hierarchy has limited demonstrated incremental benefit. On LIDP, adding masking to feature-plus-slice distillation raises AUC from 80.77% to 81.00%. Externally, adding slice distillation to feature-plus-mask raises AUC from 75.03% to 75.28%, while feature-only distillation already reaches 73.35%. These differences do not establish stable complementarity without paired uncertainty estimates. Table 4 contains only internal ablations; it does not demonstrate better external generalization than existing alternatives.

  4. Baseline comparability and clinical gains remain uncertain. Section 4.2 mixes reproduced and literature-reported results without identifying each source. Reproduce the closest alternatives on identical patient folds and clarify nested teacher/student selection. Distinguish LIDC radiologist-rating labels from LIDP pathology labels. On LIDP, sensitivity reaches 96.15% but specificity is 44.73%, versus 58.53% for CLIP-Lung. Compare sensitivity at matched specificity using development-selected thresholds before claiming a practical advantage.

References [1] Lei et al. CLIP-Lung: Textual Knowledge-Guided Lung Nodule Malignancy Prediction. MICCAI, 2023. [2] Zagoruyko and Komodakis. Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer. ICLR, 2017. [3] Kakogeorgiou et al. What to Hide from Your Students: Attention-Guided Masked Image Modeling. ECCV, 2022. [4] Yang et al. Masked Generative Distillation. ECCV, 2022.

Questions for authors
  1. Methodological novelty: What is the specific methodological advance over combining an attribute-assisted graph teacher with established feature, attention, and masking-based distillation? In particular, Eq. 15 matches global center-node features after masking. In what sense does this provide region-level knowledge transfer beyond saliency-guided augmentation?

  2. Evidence for the complete hierarchy: In external validation, feature-plus-mask distillation achieves 75.03% AUC, compared with 75.28% for the complete model. Is this difference consistent across repeated runs or supported by paired uncertainty estimates? Do any existing experiments distinguish the proposed saliency guidance from random masking with a matched masking rate?

  3. Baseline comparability: Which results in Tables 1–2 were reproduced on your patient-level splits, and which were taken from published papers? Were teacher selection, hyperparameter tuning, and threshold selection performed exclusively within the training/development partitions? Are external-validation results available for the closest existing alternatives under the same protocol?

2 comments

TO Critical Weakness 1 (Question 1):

Our contribution is a generalizable multimodal-to-unimodal distillation framework for classification tasks involving 3D images and structured data, rather than a method limited to medical imaging. We represent a 3D image as a sequence of slices and construct a multimodal teacher with a slice-attribute graph that models relationships between image slices and structured attributes. An image-only student then learns through three complementary objectives: global representation alignment, slice-level attention alignment, and saliency-guided masked distillation.

Unlike CLIP-Lung, which mainly aligns global image and attribute representations, our framework transfers knowledge at the representation, slice-attention, and region-conditioned levels. It assumes that structured attributes have observable manifestations in the images and is therefore potentially applicable to other 3D imaging tasks with complementary structured or multimodal information.

More precisely, our region-level mechanism is teacher-guided region-conditioned representation distillation. Masking locations and probabilities are jointly determined by Grad-CAM and slice-level attention from the frozen multimodal teacher. The student is trained to recover the teacher's center-node representation from the complete, unmasked input. Although Eq. (15) does not directly match local features, its gradient propagates through the student's local feature extractor and graph aggregation pathway, encouraging teacher-consistent representations under targeted regional perturbations.

This differs from conventional guided augmentation, which generally uses augmented inputs and classification labels alone. Here, the teacher determines both how the input is perturbed and the target representation to be recovered. The random-masking control further supports this design: the full model with teacher-guided saliency masking improves multiple metrics over random masking. To avoid ambiguity, we have replaced “region-level feature alignment” with “region-conditioned knowledge transfer” throughout the manuscript.

TO Critical Weakness 2 (Question 2):

We added all requested controls using the same student architecture, data splits, training budget, and evaluation protocol. The LIDC-IDRI AUCs were 95.44% for CT-only teacher logit distillation, 96.10% for attention distillation alone, 96.63% for random masking, 96.90% for multimodal teacher logit distillation, and 97.31% for simple attribute concatenation, compared with 98.15% for the full MUGAT-KD.

These results indicate that the improvement cannot be attributed solely to generic distillation, stronger supervision, or masking regularization. Instead, feature-, attention-, and region-conditioned transfer provide complementary benefits. The results are reported in Table 3 and Supplementary Tables 7 and 11, with discussion added to Section 4.3.

TO Critical Weakness 3 (Question 2):

The original external validation used a single model trained on the full LIDC-IDRI dataset and tested on LIDP, which introduced uncertainty. Supplementary Table 10 now reports the mean and sample standard deviation across five fold-specific models for both feature-plus-mask distillation and the full model. CLIP-Lung was also reproduced under the same LIDC-IDRI-to-LIDP protocol. Supplementary Table 7 additionally compares three masking-probability settings and a random-masking baseline, showing that random masking performs slightly worse than the teacher-guided strategy.

TO Critical Weakness 4 (Question 3):

We now mark literature-reported results with an asterisk. Supplementary Table 13 summarizes the dataset scale, class composition, and label definitions of all starred methods. The remaining methods in Table 1 and all methods in Table 2 were reproduced using the same patient-level splits. We also clarify that LIDC-IDRI labels are derived from radiologists’ malignancy ratings, whereas LIDP labels are based on pathological diagnoses. The teacher checkpoint was selected by validation AUC during the first 75 epochs and then frozen for student distillation.

We agree that high sensitivity at the default 0.5 threshold is insufficient to establish clinical superiority on LIDP. We therefore removed this claim. In Supplementary Table 9, thresholds are selected exclusively on the validation set using Youden’s index, and the revised manuscript explicitly discusses the sensitivity-specificity trade-off.

Thank you for the added experiments and explanations. They address some of my concerns, but the method adds only a limited amount beyond earlier work. The data splits described in the response also differ from those in the original paper. It remains unclear whether the reported results were obtained using the student model alone. I therefore keep my rating of 3 (weak acceptance).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.