ORACLE-CT: Advances in CT, Graphs & Crypto

• ORACLE-CT is a multi-domain framework incorporating a deep learning CT classification engine, a constant-time graph distance oracle, and a collision-tractable random oracle model in cryptography.
• The medical imaging instance employs organ-aware attention and scalar feature fusion to enhance interpretability and calibration in CT triage and classification tasks.
• In graph algorithms and cryptography, ORACLE-CT introduces methods that deliver constant query performance and challenge traditional security models through novel oracle designs.

ORACLE-CT refers to several distinct, high-impact frameworks across domains of medical imaging, graph algorithms, and cryptographic security. The most prominent recent instance is the “ORgan-Aware CLAssification Engine for CT” (ORACLE-CT), a supervised deep learning architecture for triage and classification of computed tomography (CT) studies, focused on organ-level localization and interpretability (Dahal et al., 19 Jan 2026). Separately, “ORACLE-CT” has denoted an approximate distance oracle with constant query time for graphs (Chechik, 2013), and the “Collision–Tractable Random Oracle Model (CT–ROM)” in cryptography (Tezuka et al., 2021). Each instantiation addresses domain-specific technical challenges using specialized algorithmic and structural innovations.

1. Organ-Aware Classification Engine for CT Imaging

ORACLE-CT (Dahal et al., 19 Jan 2026) is a modular, encoder-agnostic supervised learning pipeline for medical CT triage and classification, addressing the need for calibrated, interpretable predictions on large volumetric datasets, particularly in radiology. Standard vision-LLMs are inadequate for 3D anatomy and protocol shifts; ORACLE-CT introduces organ-level attention and feature fusion to overcome these limitations.

Model Architecture

The pipeline comprises:

• Encoder $E_\theta$: A backbone network (2D/2.5D ViT-style or 3D CNN), mapping a CT volume $X\in\mathbb R^{D\times H\times W}$ to a local feature lattice $\{u_i\in\mathbb R^d\}_{i\in\Omega}$.
• Aggregation Head $A_\phi$: Aggregates features $\{u_i\}$ into study-level logits $z\in\mathbb R^L$, supporting multiple head-modes:
  1. Global Average Pooling (GAP baseline)
  2. Global Attention (unary scorer + softmax)
  3. Organ-Masked Attention (mask-restricted softmax per organ)
  4. Organ-Masked Attention + Organ-Scalar Fusion (OSF)

Organ-Masked Attention

Given organ masks $M_o\in\{0,1\}^{D\times H\times W}$ projected onto $\Omega$, each organ group $o$ enforces spatial attention via mask-restricted softmax:

$$w_{o,i} = \frac{\exp(\tilde\alpha_{o,i}/\tau_o)m_{o,i}}{\sum_{j\in\Omega_o}\exp(\tilde\alpha_{o,j}/\tau_o)+\varepsilon}$$

where $X\in\mathbb R^{D\times H\times W}$0, yielding per-organ pooled features $X\in\mathbb R^{D\times H\times W}$1 and logits $X\in\mathbb R^{D\times H\times W}$2.

Organ-Scalar Fusion (OSF)

Scalar features—$X\in\mathbb R^{D\times H\times W}$3 (organ volume), $X\in\mathbb R^{D\times H\times W}$4 (mean Hounsfield Unit), $X\in\mathbb R^{D\times H\times W}$5 (boundary flag)—are concatenated with pooled embeddings after optional truncation down-weighting, resulting in a concatenated vector $X\in\mathbb R^{D\times H\times W}$6 input to the final per-organ MLP. This augments textural cues with morphological/density descriptors.

Training and Calibration

• Loss: Per-label BCE loss with missing label ramp-in, positive-class weighting, pos-weight clipped at 10.

• Optimization: AdamW, base LR $X\in\mathbb R^{D\times H\times W}$7, weight decay $X\in\mathbb R^{D\times H\times W}$8, head LR ×3, attention LR ×0.3, cosine schedule, early stop at 30 epochs, mixed precision, grad-clip 1.0.
• Calibration: Per-label temperature scaling and F1-optimal threshold selection on validation, frozen for test deployment.

Datasets and Empirical Results

Dataset	Model	Macro AUROC
CT-RATE (chest)	ORACLE-CT (attention+OSF)	0.86
RAD-ChestCT	ORACLE-CT (attention+OSF)	0.76
MERLIN (abdomen)	ORACLE-CT (attention+OSF)	0.85
CT-RATE	GAP baseline	0.8574
MERLIN (zero-shot VLM)	VLM baseline	0.72

Empirical improvements concentrate on size/morphology-driven labels. Organ-Masked Attention provides an AUROC increase of +0.01 over GAP baseline, OSF contributes an additional +0.01 on MERLIN. ORACLE-CT outperforms all linear-probe and zero-shot VLMs under uniform protocol (Dahal et al., 19 Jan 2026).

2. Graph Distance: ORACLE-CT Constant-Time Approximate Distance Oracle

The term ORACLE-CT also refers to an approximate distance oracle with constant query time for undirected graphs with non-negative edge weights (Chechik, 2013). This data structure extends the Thorup–Zwick and Mendel–Naor frameworks:

• Space: $X\in\mathbb R^{D\times H\times W}$9
• Query time: $\{u_i\in\mathbb R^d\}_{i\in\Omega}$0
• Stretch: $\{u_i\in\mathbb R^d\}_{i\in\Omega}$1

Data Structures and Query Algorithm

• Hierarchy: Sample vertex subsets $\{u_i\in\mathbb R^d\}_{i\in\Omega}$2.
• Storage: Closest pivots, distance layers, and “bunch” sets per vertex, merged with Mendel–Naor tables for high-stretch but quick estimates.
• Query: Two-phase process—rapid high-stretch estimate followed by constant-time layer refinement using precomputed repair indices, guaranteeing output $\{u_i\in\mathbb R^d\}_{i\in\Omega}$3 with $\{u_i\in\mathbb R^d\}_{i\in\Omega}$4.

Significance

This construction matches the optimal trade-off for stretch and space while improving query time from $\{u_i\in\mathbb R^d\}_{i\in\Omega}$5 to $\{u_i\in\mathbb R^d\}_{i\in\Omega}$6. It leverages the synergy of Thorup–Zwick’s efficient covering structures and Mendel–Naor's constant-time retrieval (Chechik, 2013).

3. ORACLE-CT in Cryptography: Collision–Tractable Random Oracle Model (CT–ROM)

In cryptography, CT–ROM is a weakened random oracle model where, in addition to the standard random oracle (RO), an adversary receives access to a collision oracle capable of returning input pairs $\{u_i\in\mathbb R^d\}_{i\in\Omega}$7 with $\{u_i\in\mathbb R^d\}_{i\in\Omega}$8 (Tezuka et al., 2021).

Model Structure

• Oracles: $\{u_i\in\mathbb R^d\}_{i\in\Omega}$9 (standard random oracle), $A_\phi$0 (returns collision pair or $A_\phi$1 if none stored).
• Security Experiments: Adversaries may alternate queries to both oracles; CT–ROM is strictly stronger than SPT–ROM or FPT–ROM, and weaker than the plain ROM.

Provable Separations

• RSA–FDH and DSA signature schemes are insecure in CT–ROM: collisions undermine unforgeability regardless of plain ROM security, with adversarial forgeries succeeding with overwhelming probability once nontrivial colliding entries arise.
• Any scheme secure in CT–ROM remains secure in the standard RO model, but insecurity in SPT–ROM or FPT–ROM does not imply insecurity in CT–ROM.

4. Implementation and Reproducibility in Organ-Aware CT

ORACLE-CT’s imaging pipeline supports diverse backbone encoders: 2.5D ViTs (DINOv3, MedSigLIP) and 3D CNNs (I3D-ResNet-121, MedNeXt-3D, CT-Net), working with mask sets derived from TotalSegmentator, post-processed for per-organ attention with customizable merges and dilations. The full experimental suite (splits, augmentations, mask rules) is available for reproducibility at the source repository.

Strict adherence to single-seed runs, early stopping, calibration, and fixed protocols allows for precise numerical reproducibility of reported metrics (Dahal et al., 19 Jan 2026).

5. Comparative Analysis and Impact

Triage and Classification

ORACLE-CT achieves state-of-the-art supervised performance on both chest and abdomen CT, exceeding contemporary vision-LLM benchmarks and providing interpretable, organ-localized outputs with auditable attention maps.

Ablation Studies

Incremental analysis shows that most performance gain derives from explicit organ localization, especially for morphologically-driven and multi-organ pathologies. Scalar cue fusion further elevates results by integrating volumetric/density information critical for certain findings.

Broader Applicability

ORACLE-CT’s encoder-agnostic and modular design enables plug-and-play experimentation across architectures and datasets, positioning it as a reference implementation for interpretable, organ-level medical AI pipelines.

This suggests that organ-aware mechanisms and explicit scalar feature fusion are core components for robust, generalizable CT study classification under protocol variation and real-world distributional shifts.

---

References:

• Organ-aware CT study-level classification: "Organ-Aware Attention Improves CT Triage and Classification" (Dahal et al., 19 Jan 2026)
• Graph distance oracles: "Approximate Distance Oracle with Constant Query Time" (Chechik, 2013)
• Collision–Tractable ROM in cryptography: "Weakened Random Oracle Models with Target Prefix" (Tezuka et al., 2021)