acceptodds
Under review as a conference paper at ICLR 2027

Multi-modal Feature Selection and Pareto-Based Model Fusion for Breast Cancer Computer-Aided Diagnosis.

Abstract

Deep learning has been widely applied to breast cancer imaging analysis, yet most existing studies optimize individual metrics such as the area under the receiver operating characteristic curve (AUC) while overlooking feature acquisition cost, feature redundancy, and prediction reliability. This paper proposes a computer-aided diagnosis method based on feature-group multi-objective optimization and Pareto model fusion. The method jointly maximizes AUC and minimizes the number of feature groups, and dynamically fuses Pareto sub-models by integrating predictive performance, feature-group count, and predictive uncertainty, yielding patient-level benign/malignant probabilities with uncertainty estimates. Experiments on clinical data and digital breast tomosynthesis (DBT) images from The Cancer Imaging Archive (TCIA) demonstrate that the proposed method reduces the number of feature groups while maintaining diagnostic performance and provides a quantitative basis for evaluating prediction reliability and stability.

Primary area:
optimization
Keywords:
computer-aided breast cancer diagnosis, multi-modality, multi-objective optimization, Monte Carlo dropout, uncertainty estimation, model fusion.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zcXpOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper combines DBT image features and structured features for benign/malignant classification. A genetic algorithm searches feature-group subsets for validation AUC and group count. Selected MLP models use Monte Carlo dropout to estimate uncertainty and are fused using validation performance, feature count, and uncertainty.

Critical strengths

None

Critical weaknesses

The claimed reduction in feature acquisition cost is not established at the ensemble level. Although each Pareto submodel may use a small feature subset, the proposed prediction uses an ensemble of 7–30 submodels. Consequently, the actual set of features that must be acquired for a patient is the union of the feature groups required by all participating submodels, which may approach the full feature set. The paper reports per-model feature counts but does not report this union or a realistic acquisition cost. This issue directly undermines one of the central motivations and objectives of the method.

The empirical evaluation is too limited to support the reported gains. The dataset contains only 256 patients, with approximately 51 samples each in the validation and test sets. The validation set is repeatedly used for the 200-generation genetic search, Pareto selection, hyperparameter tuning, candidate filtering, model ranking, and fusion-weight estimation. This creates substantial risk of selection overfitting. No nested cross-validation, repeated splits, confidence intervals, or external validation is provided. Moreover, the proposed fusion does not consistently outperform simple mean fusion or logistic regression, and the claim that it achieves the highest AUPRC under all four pool sizes is contradicted by Table 2 at m=7m=7.

Questions for authors
  1. For each ensemble size, what is the number and acquisition cost of the union of feature groups required by all participating submodels? Does the proposed fused model still reduce acquisition burden relative to the full-feature baseline?

  2. Can the authors provide repeated nested cross-validation or external-cohort results in which the complete genetic search, hyperparameter selection, filtering, and fusion-weight estimation are performed inside each training fold?

  3. Are Equations (14)–(19) derived from an existing evidential-reasoning rule? If so, what is methodologically new, and how is the combination rule justified when the submodels are strongly correlated through shared training data, mandatory image features, and overlapping clinical features?

  4. What is the exact provenance of the ultrasound and clinical variables, how were the 256 patients selected and linked to the DBT data, and how did the authors ensure that BI-RADS and radiologist-derived descriptors do not introduce target or temporal leakage?

Reviewer_qW2NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a multi modal framework for breast cancer computer-aided diagnosis that combines feature-group selection, Pareto-based candidate model selection, Monte Carlo Dropout uncertainty estimation, and preference-based model fusion. A genetic algorithm searches for feature-group subsets by jointly maximizing validation AUC and minimizing the number of selected groups. MLP sub-models are trained on candidate subsets, and their predictions are combined using weights based on validation performance, feature-group requirements, and predictive uncertainty. The authors evaluate the framework on a reported 256-patient TCIA cohort and compare it with mean fusion, a best-single-model baseline, and several classical machine-learning classifiers. Additional experiments examine uncertainty-based sample filtering and robustness to feature missingness and noise. The central claim is that the proposed framework balances diagnostic performance, feature parsimony, and prediction reliability.

Critical strengths

1.Relevant problem formulation. The paper explicitly considers diagnostic performance and feature-group parsimony rather than optimizing classification performance alone. This is a potentially useful formulation for clinical settings where collecting and maintaining every feature may be burdensome.

2.Evaluation beyond a single classification metric. The authors compare multiple model-pool sizes and fusion strategies and investigate performance under feature missingness and noise. These experiments provide an initial assessment of the proposed framework beyond its standard test-set classification results.

Critical weaknesses

1.Unclear data provenance and imaging modality. The manuscript states that it uses clinical data and digital breast tomosynthesis (DBT) images from 256 patients in TCIA. However, its feature groups include sonographic descriptors such as anechoic, hypoechoic, hyperechoic, shadowing, and enhancement. The reported cohort size and feature schema closely resemble TCIA's BREAST-LESIONS-USG collection, which contains ultrasound data from 256 patients, whereas TCIA's Breast Cancer Screening–DBT collection is a separate dataset with 5,060 patients and 22,032 scans. The manuscript does not clearly identify the exact collection or explain how the imaging and clinical variables are matched at the patient level. This is a substantive concern because the imaging modality and data correspondence are fundamental to evaluating the validity of the proposed multimodal framework. The authors should clarify the exact data source, imaging modality, inclusion and exclusion criteria, and feature provenance.

  1. Insufficient methodological novelty and empirical evidence. The framework combines established methods, including genetic algorithms, Pareto optimization, Monte Carlo Dropout, and model fusion, but its novelty beyond related work, particularly Deep-AutoMO and Adaptive-AutoMO, remains unclear. Comparisons with recent methods and ablations isolating the contributions of individual components are insufficient. Furthermore, Appendix A.7 reports 32 candidate submodels but identifies only two solutions on the Pareto front, while the experiments treat m=30 as the full model pool. The relationship between these numbers and the exclusion of two candidates require clarification. Evaluation is limited to a small single-cohort split without confidence intervals, repeated-split experiments, or external validation. Finally, Table 2 contradicts the claim that dynamic preference fusion achieves the highest AUPRC in all settings: mean fusion performs better at m=7, while logistic regression achieves higher AUC across all four settings. Overall, the evidence does not establish a robust advantage over simpler alternatives.
Questions for authors

1.Dataset and modality: Which exact TCIA collection was used, and what is its DOI? Are the input images DBT or ultrasound? Please explain whether the reported 256 patients represent the complete cohort or a filtered subset. Please also clarify the handling of the four cases described as having extensive missing data.

2.Image processing and evaluation protocol: How is each image converted into a 512-dimensional feature vector? Please specify the Transformer architecture, pretraining or training procedure, image-level and patient-level aggregation, label source, class distribution, and exact train/validation/test sample counts. Was the split performed strictly at the patient level?

3.Methodological contribution and baselines: What is the main methodological contribution beyond existing multi-objective feature selection and uncertainty-aware fusion methods, particularly Deep-AutoMO and Adaptive-AutoMO? Could the authors provide ablations that isolate the contributions of the genetic algorithm, Pareto candidate selection, and preference-based fusion?

4.Fusion, uncertainty, and numerical results: What is the relationship between the 32 candidate submodels, the two solutions described as lying on the Pareto front, and the model pools of m = 7, 15, 22, and 30? Please reconcile the statement that dynamic preference fusion achieves the highest AUPRC at every pool size with the numerical results in Table 2. Finally, how is predictive uncertainty validated beyond removing high-uncertainty samples, and can the authors report calibration or risk-coverage analyses and statistical uncertainty for the performance differences?

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

The paper proposes a breast-lesion classification pipeline that searches over combinations of imaging and structured clinical feature groups. A genetic algorithm evaluates the trade-off between validation AUC and feature-group count. Candidate subsets are used to train MLP classifiers, whose predictions are averaged over Monte Carlo dropout passes. The final fusion procedure incorporates validation performance, feature-count preferences, and predictive entropy to produce class scores and an uncertainty output.

Evaluation uses a reported 256-patient TCIA cohort with a 60/20/20 split, comparing four ensemble sizes against a best single model, mean fusion, and conventional classifiers. Additional experiments remove uncertain cases or perturb input features. The proposed fusion reaches AUC 0.8613 at seven models and its highest AUPRC, 0.8547, at 15 models. The results suggest that combining selected models can improve on the chosen single-model reference, but do not establish a consistent advantage over simpler fusion or demonstrate reduced acquisition cost for the deployed ensemble.

Critical strengths
  1. The clinical motivation is sensible. Searching over named feature groups, rather than treating every additional input as free, is a reasonable way to approach resource-constrained prediction. Retaining several accuracy–sparsity trade-offs could also be more useful than returning a single maximum-AUC configuration. The manuscript makes this design explicit.

  2. There is an effort to examine the pipeline rather than report only its final score. The staged feature-addition experiment, selection heatmaps, and comparisons across pool sizes help expose where predictive information enters the system. For example, the reported AUC rises from 0.7377 with image features alone to 0.8394 after adding morphology-related descriptors, then declines when further groups are introduced. This is a useful observation, even though its stability across splits remains unknown.

  3. The inclusion of simple competitors makes the results interpretable. Mean fusion and logistic regression are important references here, and the authors retain results in which those methods perform better. The missingness and noise experiments also broaden the evaluation beyond clean-input discrimination. These comparisons provide a useful starting point for a stronger study.

Critical weaknesses
  1. The dataset description needs to be resolved before the results can be interpreted. Section 5.1 describes DBT images from 256 TCIA patients, whereas Table 4 includes anechoic, hypoechoic, posterior enhancement, shadowing, and other ultrasound descriptors. Appendix A.5 explicitly discusses adding ultrasound findings. The exact TCIA collection is not identified.

  2. A targeted check of the public source documentation identifies BrEaST/BREAST-LESIONS-USG, a 256-patient TCIA collection containing ultrasound images and closely matching clinical and BI-RADS descriptors - not DBT. This correspondence suggests a possible modality misidentification, but does not establish which collection the authors actually used. The accession, class counts, exclusion criteria, and image-to-clinical-record linkage are necessary to resolve it. cancerimagingarchive.net

  3. The presence of BI-RADS as an input also changes the interpretation of the task. Using an existing radiologist assessment can be legitimate, but the paper should establish what the model adds beyond BI-RADS and the structured descriptors alone. Otherwise, improvements over the image-only model do not establish a benefit from the proposed fusion machinery.

  4. The uncertainty formula fails a basic disagreement test. Consider two equally weighted classifiers making opposite, fully confident predictions: $

p1=(1,0),p2=(0,1),wˉ1=wˉ2=12.p_1=(1,0),\qquad p_2=(0,1),\qquad \bar w_1=\bar w_2=\tfrac12.

$ Their individual entropies are zero. Substitution into Equations 14–19 gives $

(pbenign,pmalignant,u)=(0.5,0.5,0).(p_{\mathrm{benign}},p_{\mathrm{malignant}},u) =(0.5,0.5,0).

$ Thus, complete disagreement produces zero reported global uncertainty, although the manuscript describes this output as “conflict-induced uncertainty” and interprets larger values as lower reliability. This counterexample follows directly from the published equations; it is not an assumption about the implementation.

There is a related interpretation problem: $

∑kpk,j=1−uj,\sum_k p_{k,j}=1-u_j,

$ rather than one. These outputs could be interpreted as class-specific belief masses with residual ignorance, but they are not a normalized benign/malignant probability distribution without a further conversion. The paper needs to distinguish these objects and explain which quantity is thresholded and evaluated.

  1. Removing high-scoring uncertain cases does not resolve either issue. An ordinary confidence score may also identify easier subsets; the current experiment does not establish what Monte Carlo dropout or the proposed fusion uncertainty adds beyond such a comparator.

  2. Sparse component models do not necessarily make a low-cost ensemble. The acquisition-cost objective is defined for each candidate model, but the final system evaluates multiple models using different subsets. Unless an explicit routing or acquisition policy avoids some evaluations, deployment requires the union of their inputs: $

Gdeploy=⋃i:wˉi>0Gi.G_{\mathrm{deploy}}=\bigcup_{i:\bar w_i>0}G_i.

$ Giving a feature-heavy model a small positive weight does not eliminate the need to acquire its inputs. The paper reports per-model feature counts and then ensembles up to 30 models, without showing the resulting union-level feature requirement or acquisition cost.

Even the component-level proxy needs care: several “groups” are separate encoded categories or descriptors from the same examination, so counting them equally is not equivalent to counting independently avoidable clinical procedures. Table 4 makes this distinction visible. The paper currently demonstrates input sparsity for individual classifiers more clearly than reduced examination burden for the final system.

  1. The evidence does not yet justify the added complexity. The clean-data comparison is less favorable to the proposed method than the discussion suggests. At �4��4�, mean fusion exceeds Dynamic Preference on both AUC and AUPRC: 0.8661/0.8592 versus 0.8613/0.8484. Logistic regression has higher AUC than every reported Dynamic Preference configuration. The statement that Dynamic Preference obtains the highest AUPRC in all four settings therefore conflicts with Table 2.

The small gains at other pool sizes come from one split with roughly 50 test patients, without paired confidence intervals or repeated full-pipeline evaluation. Meanwhile, the validation set supports 200 search iterations, subsequent hyperparameter optimization, and ensemble selection. A held-out test set is appropriate, but this design leaves the stability of the selected subsets and fusion advantage largely unmeasured. Five repetitions of input perturbations do not substitute for repeating model development.

A matched-budget comparison against simpler subset selection and ordinary averaging would be more convincing than further comparisons among pool sizes selected by the proposed scoring rule.

Questions for authors
  1. Which dataset and imaging modality were actually used? Please provide the TCIA collection identifier, retained class counts, and the encoder architecture/checkpoint. Was normalization fitted exclusively on training patients? Appendix A.4 describes normalization before the split, so the fitting scope is currently unclear.

  2. Does the implementation follow Equations 11-19 exactly? How should the opposite-confident-predictions example above be interpreted, and how are the class masses converted into diagnostic probabilities? An explicit worked example would resolve more than an additional aggregate filtering table.

  3. What is the actual feature requirement of each deployed ensemble? Reporting the union of required groups, together with performance against a single model or mean ensemble constrained to the same acquisition budget, would directly test the paper’s central cost claim.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.