Technical Summary: TRACE – A Concept Bottleneck Model for Longitudinal 3D Glioblastoma Response Assessment

1. Problem Statement

Longitudinal assessment of Glioblastoma (GBM) treatment response is critical for clinical decision-making but remains challenging due to the subtlety of tumor changes over time and the complexity of clinical criteria. Clinicians rely on the Response Assessment in Neuro-Oncology (RANO) criteria (specifically RANO 2.0), which require comparing tumor measurements (volume and Sum of Products of Diameters, SPD) across baseline and follow-up MRI scans to categorize response into four classes: Complete Response (CR), Partial Response (PR), Stable Disease (SD), and Progressive Disease (PD).

Current deep learning approaches typically predict these response labels directly from imaging features. While effective in some metrics, these "black-box" models lack interpretability, making it difficult for clinicians to inspect the reasoning or correct errors. Furthermore, standard models often fail to explicitly encode the structured, causal dependencies inherent in the RANO protocol (e.g., percentage changes are deterministic functions of baseline and follow-up measurements). There is a need for an approach that aligns with clinical reasoning, offers transparency, and allows for human intervention at the concept level.

2. Methodology: The TRACE Architecture

The authors propose TRACE (Temporal Causal Reasoning with Acyclic Concept Explanation), a Concept Bottleneck Model (CBM) designed specifically for longitudinal GBM assessment. Unlike standard CBMs that treat concepts as a flat set of independent predictors, TRACE structures the bottleneck as a Directed Acyclic Graph (DAG) that mirrors the RANO decision process.

2.1. Input and Encoder

The model processes paired baseline ($t_0$) and follow-up ($t_1$) multimodal MRI scans (T1, T1ce, T2, FLAIR) along with their corresponding tumor segmentation masks.

• Siamese 3D Encoder: A shared-weight 3D ResNet-18 (MedicalNet backbone) encodes both time points separately for MRI and segmentation streams.
• Temporal Representation: The encoder outputs are concatenated with their explicit difference ($h_{t1} - h_{t0}$) to form a longitudinal feature triplet. These are fused into a joint representation ($X_{enc}$).

2.2. Structured Concept Bottleneck

The core innovation is the separation of concepts into three types within the DAG:

1. Root Concepts ($C_r$): Image-derived tumor measurements predicted by learnable "ConceptBlocks" (MLPs) from $X_{enc}$. These include baseline/follow-up volumes ($V_{t0}, V_{t1}$) and SPDs ($SPD_{t0}, SPD_{t1}$).
2. Derived Concepts ($C_d$): These are deterministic nodes computed via fixed mathematical rules from parent concepts, not learned. They include percentage changes ($\Delta V\%, \Delta SPD\%$) and binary threshold flags (e.g., $b_{PD}^{vol}$ if $\Delta V\% \ge 40\%$). This ensures mutual consistency and prevents the model from learning spurious correlations for these derived values.
3. Passthrough Concepts ($M$): Clinical metadata injected directly without image encoding, such as the scan interval ($\Delta t$) and a new-lesion flag ($b_{NL}$).

The final task head (an MLP) predicts the 4-class RANO response ($Y$) solely based on the structured bottleneck vector $(C_r, C_d, M)$, ensuring no "hidden" image features leak directly into the final classification.

2.3. Training and Intervention Consistency

To address the "train-test mismatch" where correcting a concept at test time might shift the distribution for the task head, TRACE employs Intervention-Consistency Training:

• During training, root concepts are randomly replaced with ground-truth values (intervention) with a probability $p_{int}$.
• The model is penalized via a Kullback–Leibler (KL) divergence loss ($\mathcal{L}_{IC}$) to ensure the prediction remains consistent whether the input concepts are predicted or corrected.
• The loss function combines a soft macro F1 objective (with focal weighting for class imbalance), an auxiliary binary PD-vs-rest head, and concept supervision for root nodes.

3. Key Contributions

1. First RANO-Aligned CBM: TRACE is the first Concept Bottleneck Model designed specifically for RANO-based longitudinal GBM response classification, representing clinical dependencies as a DAG rather than a flat set of predictors.
2. Deterministic Reasoning in the Bottleneck: The model encodes RANO reasoning rules (percentage changes and thresholds) as deterministic nodes. This constrains the task head to human-interpretable variables and ensures that concept-level corrections propagate logically through the reasoning chain.
3. Intervention-Consistency Training: The authors introduce a training strategy that explicitly penalizes disagreement between predictions made on predicted concepts versus corrected concepts, enabling reliable test-time intervention.
4. Evaluation on LUMIERE: The model is evaluated on the LUMIERE dataset using 5-fold patient-wise cross-validation, demonstrating performance comparable to non-interpretable baselines while offering interpretability.

4. Results

Evaluated on the LUMIERE dataset (361 longitudinal pairs from 91 patients), TRACE achieved the following:

• 4-Class Performance: A macro F1 score of 0.4769 ± 0.1229.
• Binary Performance (PD vs. Non-PD): A macro F1 score of 0.7085 ± 0.0935.
• Comparison: TRACE outperformed a Data-Efficient CBM (DCBM) baseline (0.254 macro F1) and remained within the performance range of published non-interpretable deep learning baselines (e.g., DenseNet-based models).
• Ablation Studies:
  • Removing intervention-consistency training caused a massive drop in performance (4-class F1 dropped from 0.696 to 0.354 in a specific fold), highlighting its critical role.
  • Replacing the expert DAG with a flat bipartite graph or removing the deterministic derivation chain also resulted in significant performance degradation, confirming the necessity of the structured, causal graph.
• Intervention Experiments: Correcting root concepts (e.g., tumor volume) at test time successfully propagated through the deterministic nodes to flip the final prediction (e.g., from CR to PD), validating the model's auditability.

5. Significance and Claims

The paper positions TRACE as an exploratory step toward transparent, structured concept reasoning in medical imaging. Its significance lies in:

• Auditability: It frames response assessment as a structured reasoning problem where clinicians can inspect intermediate measurements (volumes, SPDs) and intervene to correct errors, with the correction automatically updating the final diagnosis.
• Clinical Alignment: By hard-coding RANO 2.0 logic into the bottleneck, the model respects the explicit clinical criteria used in practice, unlike black-box models that learn implicit correlations.
• Design Pattern: The authors suggest this architecture serves as a design pattern for other longitudinal medical imaging tasks where decisions depend on explicit clinical criteria, temporal changes, and human-verifiable intermediate measurements.

The authors remain modest about the results, noting that performance on rare classes (CR and PR) is limited by the small dataset size and that the model requires external validation on larger, protocol-aligned datasets before broader clinical adoption. They acknowledge that while the model improves interpretability, the estimation of root concepts (especially baseline measurements) remains a challenge for future work.