Complementary Roles of Radiomics and Foundation Representations in Renal Cell Carcinoma Classification: A Comparative Study of 2D and 3D CT Encodings
Yuan Liang, Sourav Bhattacharjee, Abraham Campbell
cs.CV · cs.AI
Quick verdict
Claims radiomics+MedVAE fusion (3D gated, 82.7% AUC) beats either alone on KiTS23 RCC; plausible but rests on a 60-case test set lacking significance tests.
Summary
This comparative study evaluates whether handcrafted radiomics remains valuable when combined with pretrained foundation representations for preoperative renal cell carcinoma (RCC) subtype classification. Using 396 contrast-enhanced CT scans from the public KiTS23 challenge (multifocal cases excluded; 60/25/15 stratified train/validation/test split), the authors extract 386 IBSI-aligned PyRadiomics features reduced to 73 via reliability, variance, and correlation filtering, and derive image embeddings from partially fine-tuned 2D (top-k slice) and 3D (96-cubed patch) MedVAE autoencoders. Three fusion strategies — concatenation, cross-attention, and gated fusion — are compared against radiomics-only and image-only baselines. 3D gated fusion performs best (82.7% test AUC), ahead of the best 2D fusion (79.6%), the radiomics ML baseline (74.4%), and the single-modality MedVAE branches (59.5% 2D, 63.2% 3D). Ablations of the best model yield 60.4% (image-only) and 52.9% (radiomics-only), supporting complementarity; permutation importance ranks GLDM dependence entropy first. The main limitations are a small test set (~60 cases, apparently ~18 non-ccRCC), the complete absence of confidence intervals or significance tests behind headline deltas of 2-3 AUC points, unstated label provenance for histologic subtype within a segmentation challenge dataset, several unspecified implementation details (e.g., the ML baseline classifier), and no external validation or code availability statement.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fusing handcrafted radiomics with MedVAE representations consistently improves ccRCC vs non-ccRCC discrimination over image-only MedVAE branches on KiTS23 (e.g., 3D gated fusion 82.7% AUC vs 63.2% image-only).
- The best model is 3D gated fusion at 82.7% test AUC, exceeding the best 2D fusion model (79.6%) and the radiomics ML baseline (74.4%).
- Handcrafted radiomics remains relevant in the foundation-model era: the radiomics baseline (74.4% AUC) outperforms both single-modality MedVAE branches (59.5% 2D, 63.2% 3D).
- Ablations of the best model (image-only 60.4%, radiomics-only 52.9% vs full fusion 82.7%) show the learned interaction of both modalities drives performance.
- Permutation importance within the fusion model (top feature: original_gldm_DependenceEntropy) provides decision-centric interpretability that reflects the internal decision pathway.
Strengths
Fair, unified comparison protocol
All eight model variants share one preprocessing pipeline (§2.1: RAS, 1 mm isotropic, HU [-150, 200]), one stratified case-level 60/25/15 split, training-set-only z-scoring, and a validation-set Youden threshold applied unchanged to test (§3.1) — leakage controls are sound and comparisons are apples-to-apples.
Systematic factorial design with honest baselines
2 modality branches x 3 fusion mechanisms (Table 1), plus a radiomics baseline and ablations of the best model (Table 2: 82.7 vs 60.4 vs 52.9). Notably, the authors report unflattering results plainly: both image-only MedVAE branches (59.5%, 63.2% AUC) lose to plain radiomics (74.4%).
Careful radiomics standardization
Decision-centric interpretability framing
Weaknesses
Statistically underpowered evaluation
Unexplained label provenance in a segmentation dataset
Missing implementation details and code
Uncontrolled contrast-phase confounding
Asymmetric gated fusion design
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether handcrafted radiomics still adds value once pretrained "foundation" image representations are available, and whether 2D or 3D MedVAE encoders serve RCC CT classification better. On 396 KiTS23 contrast-enhanced CTs (multifocal cases excluded), it compares a radiomics ML baseline, 2D and 3D MedVAE branches, and three fusion mechanisms (concatenation, cross-attention, gated fusion) for binary ccRCC vs non-ccRCC classification. Central claims (Abstract, §3.2-3.4): fusion consistently improves over image-only branches; 3D gated fusion is best at 82.7% test AUC vs the best 2D fusion at 79.6%, the radiomics baseline at 74.4%, and image-only branches at 59.5%/63.2%; ablations (Table 2: 60.4% image-only, 52.9% radiomics-only) prove complementarity; and permutation importance inside the fusion model yields decision-centric interpretability, with original_gldm_DependenceEntropy ranked first.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the best result reported in this RCC CT classification paper?
The 3D gated fusion model reaches 82.7% AUC for distinguishing clear-cell from non-clear-cell renal cell carcinoma on a ~60-case test split of the public KiTS23 CT dataset, ahead of the best 2D fusion model (79.6%) and a radiomics-only baseline (74.4%).
Did the MedVAE foundation model beat radiomics features?
No. Used alone, the 2D and 3D MedVAE branches reached only 59.5% and 63.2% AUC — below the 74.4% radiomics baseline. Radiomics helped only in combination, where fusing it with MedVAE features gave the best results.
What dataset was used and how was it split?
396 contrast-enhanced CT scans from the public KiTS23 kidney-tumor segmentation challenge, with multifocal cases excluded, split 60% train / 25% validation / 15% test at the case level with stratification to avoid leakage.
Is code or data available for this paper?
The paper uses the public KiTS23 dataset and standard tools (MONAI, PyRadiomics, MedVAE), but the text contains no code repository link or explicit code availability statement, and several implementation details (e.g., the baseline classifier) are unreported.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2's top-10 feature values are not extractable from text; only the top-ranked feature (original_gldm_DependenceEntropy) is named in the body.
- Test-set size (~60) and non-ccRCC count (~18) are inferred from the stated 60/25/15 split and the 5.6% specificity granularity in Table 1, not stated explicitly by the authors.
- No code or data availability statement appears in the extracted text; codeAvailable=false reflects its absence, not an explicit statement that code is closed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.