TumorCoT: CT Tumor Analysis Benchmark

• TumorCoT is a large-scale multimodal benchmark for clinical tumor analysis, featuring 1.5M chain-of-thought labeled VQA samples paired with 3D CT scans.
• It supports detailed tasks such as localization, lesion attribute evaluation, TNM prediction, and structured report generation via stepwise reasoning.
• The dataset utilizes rigorous CT preprocessing and a knowledge-graph guided annotation pipeline to ensure clinical traceability and high performance metrics.

Searching arXiv for papers relevant to TumorCoT and TumorChain. {"query":"TumorCoT TumorChain arXiv 2026 clinical tumor analysis multimodal chain-of-thought", "max_results": 10} TumorCoT is a large-scale multimodal benchmark for clinical tumor analysis that operationalizes a reasoning trajectory from findings to impression to pathology-level prediction. It was introduced together with the TumorChain framework as a dataset of 1.5M CoT-labeled VQA instructions paired with 3D CT scans, with step-aligned rationales and cross-modal alignments designed to support evaluation of both answer accuracy and reasoning consistency in tumor-centric radiologic reasoning (Li et al., 6 Mar 2026).

1. Dataset definition and task scope

TumorCoT-1.5M comprises 41,059 3D CT studies, 10,708 radiology reports, partial pathology reports, and 1,497,818 multimodal VQA instruction samples centered on five digestive organs: liver, pancreas, stomach, colon, and esophagus. The dataset uses a strict patient-level split (9:1 train:test) to ensure no leakage, and it includes both multiple-choice and open-ended QA, organized by organ substructures and tumor grades (Li et al., 6 Mar 2026).

The benchmark is structured around four core task families with stepwise CoT rationales:

Task family	Scope
Localization	Organ/tumor positions, segment localization
Lesion attributes	Shape, boundary, density, count, and others
TNM prediction	Tumor, node, metastasis
CoT report generation	Finding → impression → pathology

This organization is consequential because TumorCoT is not limited to end-to-end report generation. It spans low-level localization, lesion characterization, staging-related prediction, and long-form reasoning. The dataset therefore formalizes tumor analysis as a multimodal reasoning problem rather than as a single classification or captioning task (Li et al., 6 Mar 2026).

2. Data construction, preprocessing, and annotation pipeline

The 3D CT preprocessing pipeline follows a fixed radiologic standardization procedure: soft-tissue window extraction, voxel-intensity normalization, cropping and zero-padding to handle shape heterogeneity without geometric distortion, and final resizing to $256 \times 256 \times 32$ for the M3D encoder (Li et al., 6 Mar 2026).

TumorCoT’s rationale annotation is produced by a multi-agent, knowledge-graph–guided data engine that converts radiology and pathology reports into step-aligned CoT chains. The pipeline includes several explicit components. TotalSegmentator generates 117 organ masks, which are merged into 56 and then refined by radiologists for the 5 digestive organs. Qwen3-235B-A22B is used as a structured feature extractor to standardize terminology with RadLex and extract structured medical entities across findings, impressions, histopathology, and TNM. A triplet-based knowledge graph integrates authoritative guidelines, textbooks, and expert-labeled cases to constrain prompts and enforce traceable, guideline-consistent logic (Li et al., 6 Mar 2026).

Reasoning is then represented as subject–relation–object triplets across three linked chains: Finding Chain (FC), Impression Chain (IC), and Long Reasoning Chain (LRC). This is a defining property of TumorCoT: rationale supervision is not merely free text, but a structured chain representation aligned to clinical semantics and later reused for evaluation (Li et al., 6 Mar 2026).

Clinical validation is reported through an expert audit of 5,000 sampled instances, yielding a 95.88% usability rate; among usable samples, 97.85% are high-quality, supporting the claim that the generated CoT chains are clinically plausible and logically consistent (Li et al., 6 Mar 2026).

3. Reasoning representation and evaluation protocol

TumorCoT explicitly models the clinical sequence finding → impression → pathology-level prediction. Its chain representation is evaluated with TumorChain-Eval, which aggregates performance on FC, IC, and LRC rather than relying only on surface-form textual overlap (Li et al., 6 Mar 2026).

The paper defines the aggregate score as

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$

with $W_{FC}+W_{IC}+W_{LRC}=1$, and the reported weights are $W_{FC}=0.3$, $W_{IC}=0.3$, and $W_{LRC}=0.4$ (Li et al., 6 Mar 2026).

This evaluation design matters because TumorCoT is intended to score not only whether a system reaches the correct endpoint, but whether it follows a clinically coherent intermediate chain. A plausible implication is that the benchmark is optimized for traceability and auditability rather than for answer-only performance.

4. TumorChain as the reference modeling framework

TumorCoT was introduced alongside TumorChain, a multimodal interleaved reasoning framework that couples a 3D CT scan encoder, an MLP projector, an organ segmentation expert, an auxiliary organ-level abnormality classifier, and a clinical text encoder based on Qwen2.5-VL-3B/7B (Li et al., 6 Mar 2026).

The visual backbone is M3D, which ingests full CT volumes $V_{ct} \in \mathbb{R}^{H \times W \times D}$. It produces volumetric visual tokens $\tau_v = E_v(V_{ct})$, which are mapped by a 2-layer MLP (Linear–ReLU–Linear) into the LLM embedding space as global tokens $\tau_g = P(\tau_v)$. TotalSegmentator produces multi-organ masks $M_{organ} = Seg(V_{ct})$, and task-specific ROI masks are selected by matching the task prompt to organ names, yielding local tokens through

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$0

An auxiliary classifier $$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$1 predicts normal/abnormal logits per organ region from $$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$2, with the stated purpose of enhancing ROI discrimination and stabilizing alignment between local visual evidence and text attention during LLM inference (Li et al., 6 Mar 2026).

The core inference mechanism is organ-guided Iterative Interleaved Reasoning (IIR). The LLM input is an interleaved sequence of global tokens, task prompt, and organ-level local tokens:

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$3

The first pass performs global reasoning, producing $$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$4. The model then performs self-reflection and organ localization, extracts the next ROI tokens, augments the prompt, and re-injects the initial answer and localized evidence. Iteration continues until no new ROIs emerge and the CoT converges on study-level impressions and pathology predictions (Li et al., 6 Mar 2026).

Training uses a CoT generation objective

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$5

together with an auxiliary organ abnormality classification term

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$6

combined as

$$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$7

Two technical clarifications are explicit in the paper. First, TumorChain does not introduce a separate contrastive objective such as InfoNCE; cross-modal alignment is enforced implicitly through the projector, ROI extraction, and joint optimization. Second, it does not train a 3D bounding-box detector or optimize bounding-box regression losses. Lesion localization is instead supervised and evaluated through CoT token generation accuracy and organ-level classification loss (Li et al., 6 Mar 2026).

5. Reported performance, generalization, and ablations

On the TumorCoT test split, TumorChain-7B achieves state-of-the-art average accuracy of 84.41% across all subtask families. Reported localization performance is 99.97% for Position: Organ, 97.57% for Position: Tumor, and 86.88% for Seg. Loc. Lesion-attribute scores are 82.28% for Shape, 84.52% for Boundary, 85.05% for Density, 86.20% for Count, and 86.57% for Others. TNM prediction reaches 88.83% for Tumor, 61.63% for Node, and 71.07% for Metastasis. The CoT-Report score is 82.36% (Li et al., 6 Mar 2026).

For chain-quality evaluation, TumorChain-7B reports $$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$8, $$CoT_e = W_{FC} \cdot \frac{1}{N}\sum_{i=1}^N S_{FC}^i + W_{IC} \cdot \frac{1}{N}\sum_{i=1}^N S_{IC}^i + W_{LRC} \cdot \frac{1}{N}\sum_{i=1}^N S_{LRC}^i,$$9, $W_{FC}+W_{IC}+W_{LRC}=1$0, and $W_{FC}+W_{IC}+W_{LRC}=1$1. The paper states that it outperforms most commercial and open-source baselines while being slightly behind GPT-5-mini on $W_{FC}+W_{IC}+W_{LRC}=1$2, which the authors interpret as evidence that higher-order reasoning remains improvable (Li et al., 6 Mar 2026).

Appendix-level organ-specific CoT report scores for TumorChain-7B are Liver 83.45, Pancreas 78.43, Colon 68.65, Stomach 93.45, and Esophagus 88.71, with average 82.54. On DeepTumorVQA, without any finetuning, TumorChain-7B achieves Recognition 73.30, Visual reasoning 53.31, Medical reasoning 45.93, and Average 57.51, exceeding the second-best reasoning baseline by +14.84. For overlapping organs, the reported generalization scores are Liver 55.41, Pancreas 62.45, Colon 84.33, and Average 67.40 (Li et al., 6 Mar 2026).

Ablations attribute performance to both CoT supervision and interleaved reasoning. Removing both CoT and IIR reduces average accuracy to 79.90%. Removing only CoT yields 82.45%, and removing only IIR yields 80.34%. The full model reaches 84.41%, a +5.64% gain over the weakest ablation. The paper also reports that adding IIR and the segmentation expert introduces ~2.51 seconds/sample extra inference time but boosts accuracy by ~4%. Sensitivity analysis shows the best result at $W_{FC}+W_{IC}+W_{LRC}=1$3; both $W_{FC}+W_{IC}+W_{LRC}=1$4 and $W_{FC}+W_{IC}+W_{LRC}=1$5 degrade performance (Li et al., 6 Mar 2026).

These results position TumorCoT not merely as a dataset for finetuning, but as a benchmark that can expose differences between answer correctness, chain fidelity, and cross-dataset transfer.

6. Limitations, clinical significance, and terminological boundaries

The principal limitations stated for the TumorCoT ecosystem are architectural rather than curatorial. TumorChain has no end-to-end lesion detector; it relies on a segmentation expert for ROI masks and a binary organ-level abnormality head. Reported qualitative error modes include miss-detections at organ boundaries, false positives due to overlapping CT appearances such as pancreatitis vs. malignancy, and misattribution of primary organ due to mass effect/compression. The authors also identify a gap between TumorChain and GPT-5-mini on $W_{FC}+W_{IC}+W_{LRC}=1$6 as evidence that long-chain clinical reasoning remains an open problem (Li et al., 6 Mar 2026).

The stated clinical implications are traceable reasoning, topology-aware analysis, and more clinically relevant benchmarking. By interleaving global context with organ-level ROI evidence, TumorChain is described as emulating reasoning across adjacent and downstream structures such as the liver–spleen–portal system, which is directly relevant to staging and metastasis assessment. TumorChain-Eval is presented as a clinically oriented alternative to surface-text metrics because it scores FC, IC, and LRC explicitly (Li et al., 6 Mar 2026).

Future directions proposed in the paper include expanding beyond the current five digestive organs, incorporating richer clinical signals such as longitudinal scans, multi-modality imaging, and lab results, integrating trainable ROI detectors or segmenters, introducing stronger knowledge integration in training or inference, and exploring reinforcement learning or multi-agent self-check strategies to improve LRC performance under difficult cases (Li et al., 6 Mar 2026).

A recurrent source of confusion is nomenclature. TumorCoT is the dataset and benchmark defined in the TumorChain clinical tumor analysis work; it should not be conflated with the “Chain-of-Cancer” multimodal survival prediction framework, which uses pathology, genomics, methylation, and language prompts for hazard-based survival modeling (Zhou et al., 28 May 2025). It is also distinct from the older cellular-automaton use of “TumorChain” to describe one-cell-wide invasive branches emerging in heterogeneous microenvironments, where the term refers to a mechanistic morphology rather than a multimodal reasoning dataset (Jiao et al., 2011). This suggests that, within recent oncology AI literature, “TumorCoT” has a specific and comparatively narrow meaning: a CT-grounded, CoT-supervised benchmark for traceable tumor analysis.