Radiomics-Conditioned Modulation of RenalCLIP Features for Clear Cell Renal Cell Carcinoma Classification
Yuan Liang, Sourav Bhattacharjee, Abraham Campbell
cs.CV · cs.AI
Quick verdict
Claims radiomics-FiLM tops RenalCLIP fusion for ccRCC CT classification (AUC 0.804/0.854); margins over simpler fusion and a strong SVC are small and untested.
Summary
The paper proposes a radiomics-conditioned residual FiLM framework for classifying clear cell renal cell carcinoma (ccRCC) from CT. Frozen RenalCLIP image features (mask-weighted pooled, 256-d) are modulated by 73 harmonised PyRadiomics descriptors through a conditioning network, while the projected radiomics vector also enters the classifier directly; zero-initialised conditioning starts at the average of the two representations. On 396 KiTS23 patients (237/99/60 split) and an external cohort of 111 patients from RCC-AID and TCGA/AIMI, FiLM attains internal AUC 0.804 ± 0.008 and external AUC 0.854, the highest mean among tested RenalCLIP fusion strategies (concatenation, radiomics/joint gating, global affine). Fusion lifts image-only RenalCLIP by ~9.8 percentage points internally. However, margins over global affine modulation (0.799 internal) and concatenation (0.848 external) are small, no significance tests are reported, and a radiomics SVC alone reaches 0.859 external AUC. Inference-time ablations show removing the direct residual or modulation each costs under one AUC point, indicating overlapping information routes; permutation importance is dominated by LoG texture descriptors. Limitations include single-run convolutional/SVC references, 60/111-patient test sets with wide CIs, no code availability statement, and cautious interpretation of frozen-versus-end-to-end comparisons. A careful, honest study whose headline architecture contribution is only weakly differentiated from simpler fusion.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
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
- Radiomics-conditioned residual FiLM attains the highest mean AUC among evaluated RenalCLIP fusion strategies in both cohorts (0.804 internal, 0.854 external).
- Adding radiomics to frozen RenalCLIP features improves internal AUC by ~9.8 percentage points over the image-only classifier (0.7063 → 0.8038).
- Both conditional modulation and the direct radiomics residual contribute, but with heavily overlapping information — removing either costs less than 1 AUC point at inference.
- Permutation importance indicates tumour texture descriptors (LoG GLCM IMC2, GLDM dependence entropy, GLSZM zone entropy) dominate model reliance, beyond size or a single intensity summary.
Strengths
Leakage-controlled evaluation design
Fixed patient-level split (237/99/60) with stratified ~70% ccRCC prevalence, train-fitted imputation/standardisation applied unchanged to validation/test/external, model selection on validation AUC only, and independent external validation on 111 patients (§2.1, §2.5).
Standards-compliant radiomics pipeline
PyRadiomics with IBSI definitions (§2.2): 386→73 features via ICC(3,1) ≥ 0.75 stability, MAD percentile cut, |ρ|>0.95 correlation clustering (average linkage, 0.2 cut); LoG σ∈{1,2,3} mm and 25 HU bins all specified.
Unusually honest interpretation
Thoughtful mechanism analyses
Weaknesses
No significance testing for the central comparison
The titular mechanism contributes almost nothing
Asymmetric baseline treatment
Very small evaluation sets
No code/split release; external heterogeneity unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This MICCAI-style manuscript (arXiv:2609.26492) addresses preoperative ccRCC vs non-ccRCC classification on CT in a small labelled cohort. The claimed contribution is a radiomics-conditioned residual FiLM framework: 73 IBSI-standardised PyRadiomics descriptors condition a tanh-bounded affine transformation (γ, β) of frozen RenalCLIP features while the projected radiomics vector enters the fused representation directly (Eq. 3, zero-initialised to (i+r)/2). Verifiable claims from Tables 1–2: FiLM attains the highest mean AUC among tested RenalCLIP fusion strategies (0.8038±0.0081 internal on 60 test patients; 0.8535 external on 111 patients); fusion lifts image-only RenalCLIP (0.7063 internal) by ~9.8pp; pathway ablations and permutation importance (§3.3–3.4, Fig. 2) characterise mechanism. The abstract's framing — 'identify FiLM as an effective approach' — overstates what the data support relative to the paper's own comparators; §3.5 itself is more careful.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 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 AUC does the proposed FiLM model achieve?
0.804 (mean ± 0.008 over three runs) on the 60-patient internal KiTS23 test set and 0.854 on the 111-patient external cohort — the highest mean among the tested RenalCLIP fusion strategies. Note a simple radiomics support-vector classifier already reaches 0.766 internal and 0.859 external AUC, so the added complexity is not clearly better on external data.
What is RenalCLIP?
A disease-specific vision-language foundation model for kidney cancer, pretrained on large independent institutional cohorts in China. This paper uses its frozen 3D ResNet18 image encoder to produce a 256-dimensional tumour representation from CT, without any fine-tuning of the encoder.
What data was the model trained and tested on?
396 patients from the public KiTS23 challenge as the internal cohort (split 237 train / 99 validation / 60 test, about 70% clear cell RCC) and 111 external patients assembled from the public RCC-AID dataset and TCGA cases with BAMF/AIMI segmentations (76.6% clear cell RCC). All source datasets are publicly available.
Is code available for this paper?
The paper does not state that code, trained models, or the exact data splits are released. The ingredients are public (KiTS23, RCC-AID, Zenodo segmentations, the public RenalCLIP encoder, PyRadiomics), but reproducing the reported numbers would require rebuilding the preprocessing, feature selection, and training pipeline from the text alone.
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.
- Figures 1-2 are only available as text descriptions; the architecture diagram (Fig. 1) and the permutation-importance plot (Fig. 2) could not be visually inspected.
- No supplementary material is included in the extracted text, so details beyond the main text (exact hyperparameters, split files, external exclusion counts) are unverifiable.
- arXiv v1 preprint; peer-review status unknown at time of review.
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.