Complementary Roles of Radiomics and Foundation Representations in Renal Cell Carcinoma Classification: A Comparative Study of 2D and 3D CT Encodings

1 Introduction

• Histologic subtype is associated with prognosis and treatment decisions, but CT interpretation faces inter-reader variability and difficulty quantifying subtle intratumoral heterogeneity.Source: §1 · p. 2
• Radiomics quantifies tumour phenotype through engineered descriptors, while MedVAE offers compact pretrained 2D and 3D representations for downstream classification.Source: §1 · p. 2
• The introduction identifies open questions about radiomics’ added value, 2D versus 3D encoders, and clinically meaningful interpretability of fusion decisions.Source: §1 · pp. 2–3
• The study combines handcrafted radiomics with MedVAE representations under consistent preprocessing and compares 2D slice-based with 3D volumetric encoders.Source: §1 · p. 3
• It evaluates concatenation, cross-attention, and gated fusion to examine radiomics–foundation interaction and assess radiomics’ added value beyond image representations.Source: §1 · p. 3
• The authors motivate internal contribution analysis of radiomics descriptors as a decision-centric alternative to relying only on post hoc explanations.Source: §1 · p. 3

2 Methods

• The framework compares radiomics, 2D and 3D MedVAE representations, and their fusion for RCC subtype classification, with radiomics importance analyzed within the fusion decision pathway.The study uses a unified preprocessing protocol.Source: §2 · p. 3, Figure 1 · p. 4

2.1 Dataset and preprocessing

• KiTS23 contributed 396 preoperative CT cases; multifocal cases were excluded, volumes were standardized to 1-mm RAS, and tumour-only radiomics and expanded-margin deep-learning ROIs were formed.When multiple tumour annotations were available, a consensus mask was generated with STAPLE.Source: §2.1 · pp. 4–5

2.2 Radiomics feature extraction and normalisation

• Radiomics extraction produced 386 initial features, then reliability, variability, correlation, and clustering filters reduced the selected vector to 73 features.Fusion models standardize selected features using training-set statistics; the standalone radiomics baseline uses extracted features without that standardization.Source: §2.2 · p. 5

2.3 MedVAE foundation representations

• The MedVAE branches use top-k tumour-bearing slices for 2D encoding or a tumour-centred volume for 3D encoding; partial fine-tuning performed best among tested settings.Unless otherwise stated, reported MedVAE results use partial fine-tuning.Source: §2.3 · p. 5, 2D MedVAE pathway: slice selection and… · p. 6, 3D MedVAE pathway: tumour-centred… · p. 6

2.4 Fusion models

• Projected radiomics and MedVAE embeddings were integrated by concatenation, image-query/radiomics-key-value cross-attention, or radiomics-conditioned channel-wise gating.The attended output is fused with the image embedding for classification in both pathways.Source: §2.4 · p. 6, Concatenation · p. 7, Cross-attention fusion · p. 7

2.5 Fusion-based interpretability analysis

• Radiomics contributions within fusion models were ranked by the mean AUC decrease after permuting each feature across evaluation cases, repeated five times per feature.The analysis focuses on model decision-pathway contributions rather than post hoc saliency maps.Source: §2.5 · p. 8

3.1 Experimental setup

• Evaluation used weighted cross-entropy, stratified 60/25/15 case splits, and validation-AUC model selection, with augmentation restricted to training-time deep-learning inputs.Source: §3.1 · p. 8

3.2 Overall performance comparison

• 82.7% AUC for 3D gated fusion exceeded 79.6% for the best 2D gated fusion and 74.4% for radiomics; fusion also beat image-only branches.Source: §3.2 · pp. 8–9, Table 1 · p. 9

3.3 Ablation analysis

• 82.7% AUC for full 3D gated fusion compared with 60.4% for image-only and 52.9% for radiomics-only, supporting contributions from both branches.Source: §3.3 · pp. 9–10, Table 2 · p. 10

3.4 Fusion-based interpretability results

• Permutation importance ranked original_gldm_DependenceEntropy highest; texture families, LoG-filtered descriptors, and first-order features informed decision-centric inspection.The figure ranks features by mean AUC decrease after feature-wise permutation, with error bars showing standard deviation across repeated permutations.Source: Figure 2 · p. 10, §3.4 · p. 11

• Radiomics remains valuable alongside MedVAE representations, and integrating the modalities was more effective than relying on either branch alone.Source: §4 · pp. 11–12
• The stronger 3D performance is consistent with RCC heterogeneity distributed across tumour volume, which 2D slice selection and pooling compress.Source: §4 · p. 11
• Ranking radiomics descriptors provides a structured view of model behaviour that can connect predictions to recognisable tumour properties for preoperative decision support.The discussion also suggests similar fusion designs may be relevant to other imaging tasks requiring discrimination and communicable model behaviour.Source: §4 · pp. 11–12

• Radiomics remains clinically relevant with pretrained image representations, and the strongest performance came from fusing the two.Source: §5 · p. 12
• The best overall model used the 3D setting, supporting preservation of volumetric tumour structure for RCC characterisation.Source: §5 · p. 12
• The framework identifies contributing radiomics descriptors for structured interpretability, and the fusion paradigm may transfer to other medical imaging tasks.Source: §5 · p. 12