Dr.V: Diverse AI Applications in Medicine

• Dr.V is an umbrella term for context-dependent AI systems enabling structured, doctor-like reasoning in medical imaging, video analysis, and clinical screening.
• Key implementations include DrVD-Bench for image-based diagnosis, Dr.V-Agent for video hallucination detection, Virtual Doctor for diabetes screening, Dr.VAE for drug response prediction, and Doctor2Vec for clinical trial recruitment.
• These systems share a staged inference design that decomposes complex clinical tasks into interpretable, evidence-linked intermediate representations.

In arXiv-linked research usage, “Dr.V” is not a single canonical system but an overloaded designation that refers to several distinct constructs: a goal of “doctor-like visual reasoning” operationalized by DrVD-Bench in medical image diagnosis, a hierarchical perception–temporal–cognition framework for diagnosing hallucinations in large video models, an interactive “Virtual Doctor” for non-invasive diabetes screening, and model families whose names begin with “Dr.V,” including Dr.VAE for drug response prediction and the Doctor2Vec representation-learning framework as labeled in the supplied account (Zhou et al., 30 May 2025, Luo et al., 15 Sep 2025, Spänig et al., 2019, Rampasek et al., 2017, Biswal et al., 2019). The term therefore functions less as a stable technical label than as a recurring abbreviation across biomedical AI, multimodal reasoning, and clinical machine learning.

1. Terminological scope and disambiguation

The most precise interpretation of “Dr.V” depends on the surrounding domain. In medical vision-language evaluation, the term denotes “doctor-like visual reasoning” rather than a standalone model; the concrete artifact is DrVD-Bench, the “Doctor-like Visual Diagnosis Benchmark” (Zhou et al., 30 May 2025). In video understanding, “Dr.V” is the explicit name of a hierarchical framework consisting of Dr.V-Bench and Dr.V-Agent for diagnosing hallucinations through fine-grained spatial-temporal grounding (Luo et al., 15 Sep 2025). In interactive clinical screening, the “Virtual Doctor” is described as an autonomous cabin that conducts a standardized anamnesis and estimates the probability of type 2 diabetes mellitus from non-invasive inputs (Spänig et al., 2019). In drug-response modeling, Dr.VAE is a semi-supervised variational autoencoder that learns pre- and post-treatment latent gene states together with response labels (Rampasek et al., 2017). The supplied account also labels Doctor2Vec as “Dr.V,” using the abbreviation for a doctor-vector representation-learning system for clinical trial recruitment (Biswal et al., 2019).

Usage of “Dr.V”	Domain	Concrete artifact
Doctor-like visual reasoning	Medical image diagnosis	DrVD-Bench
Hierarchical video hallucination diagnosis	Large video models	Dr.V-Bench and Dr.V-Agent
Virtual Doctor	Interactive diabetes screening	Autonomous cabin with DNN risk model
Dr.VAE	Drug response prediction	Semi-supervised VAE
Doctor2Vec (“Dr.V” in supplied account)	Clinical trial recruitment	Trial-conditioned doctor embedding model

This multiplicity creates a recurrent source of confusion. A common misconception is that “Dr.V” names one unified research program. The supplied records instead indicate separate lines of work with different objectives, datasets, and evaluation protocols.

2. Doctor-like visual reasoning in medical image diagnosis

In the medical-imaging context, “Dr.V” is the target capability of doctor-like visual reasoning, and DrVD-Bench is the benchmark designed to measure it (Zhou et al., 30 May 2025). The benchmark mirrors a clinical workflow extending from modality and view recognition through anatomy and lesion understanding to diagnosis and report generation. It is organized as a five-level reasoning cascade: Level 0 Image Quality, Level 1 Basic Information, Level 2 Anatomy-Level, Level 3 Lesion-Level, and Level 4 Clinical Interpretation.

DrVD-Bench contains three modules aligned to that workflow. Module 1, Visual Evidence Comprehension, comprises 4,480 expert-curated image–question pairs across 16 tasks spanning CT, MRI, ultrasound, radiography, and pathology. Module 2, Reasoning Trajectory Assessment, uses 487 images to generate 3,321 question–answer turns across Independent QA, Joint QA, and Multi-turn QA. Module 3, Report Generation Evaluation, contains 475 open-ended prompts with clinically detailed references curated from PubMedVision and PathMMU (Zhou et al., 30 May 2025).

The benchmark scale is explicitly reported as 7,789 image-question pairs across 20 task types, 17 diagnosis categories, 38 organ/tissue classes, and 27 lesion classes. For multiple-choice tasks, the metric is accuracy, with

$\text{Accuracy} = \frac{C}{T}.$

For report generation, the paper uses a normalized BERTScore based on PubMedBERT,

$$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$

No BLEU, ROUGE, CIDEr, SPICE, or statistical tests are used; results are averaged over five runs (Zhou et al., 30 May 2025).

The benchmarked models number 19 in total, including proprietary systems such as GPT-4o, GPT-o1, GPT-o3, Gemini 2.5 Pro, Grok-3, Doubao1.5-VisionPro, and Claude 3.7 Sonnet; open-source systems such as Qwen2.5-VL, LLaVA-1.6-34B, GLM-4V-9B, Janus-Pro-7B, and Phi-4 14B; and medical-specific systems including HuatuoGPT-Vision-34B, HealthGPT-L14B, RadFM-14B, and LLaVA-Med-7B (Zhou et al., 30 May 2025).

The headline empirical finding is a sharp decline in accuracy as reasoning complexity increases. For CT in Module 1, GPT-o3 is reported at 94% on Basic Information, 70% on Anatomy, 28% on Lesion, and 52% on Diagnosis. The benchmark also highlights “overdiagnosis without understanding”: on CT lesion erasure detection, GPT-o3 attains 19% and Gemini 2.5 Pro 16%, yet their CT diagnosis scores are 52% and 71%, respectively (Zhou et al., 30 May 2025). This indicates that correct final labels can occur without correct lesion grounding. A plausible implication is that benchmarking only end diagnoses can overestimate clinical reasoning competence.

The reasoning-trajectory results further distinguish static from stateful reasoning. Joint QA is consistently best, Independent QA is intermediate, and Multi-turn QA is worst. This suggests that current VLMs benefit from comprehensive static context but remain weak at maintaining clinically coherent stepwise reasoning across turns (Zhou et al., 30 May 2025).

3. Hierarchical video hallucination diagnosis

In video understanding, “Dr.V” is the formal name of a framework for diagnosing hallucinations in large video models by fine-grained spatial-temporal grounding (Luo et al., 15 Sep 2025). The framework consists of Dr.V-Bench, a benchmark with 10,000 instances from 4,974 videos, and Dr.V-Agent, a training-free, tool-augmented diagnostic agent. The system is designed around a three-level taxonomy of hallucinations: perceptive, temporal, and cognitive.

Perceptive hallucinations cover static spatial properties, including object recognition, number, color, location, static spatial relations, and OCR. Temporal hallucinations cover action recognition, dynamic attributes, dynamic relations, and sequence understanding. Cognitive hallucinations cover factual prediction, counterfactual prediction, context-based explanation, and knowledge-based explanation (Luo et al., 15 Sep 2025). Diagnosis is framed as consistency checking between a target LVM answer $A$, a video $V$, and a textual query $T$.

Dr.V-Agent uses a structured six-step pipeline. It first classifies hallucination type with GPT-4o by extracting objects $O=\{o_1,\dots,o_n\}$, events $E=\{e_1,\dots,e_m\}$, and causal claims $C=\{c_1,\dots,c_k\}$, achieving 99.6% accuracy on a 1,000-sample validation. It then performs perceptive-level checking with Grounded SAM 2 and YOLO-World, temporal-level checking with CG-STVG and Grounded-VideoLLM, and cognitive-level checking with InternVL2 and Qwen2-VL. DeepSeek-R1 reasons over the assembled evidence to identify inconsistencies, and the final feedback is structured as

$\mathcal{F} = (\mathcal{A}, \mathcal{R}),$

where $\mathcal{A}$ contains extracted spatial-temporal-causal evidence and $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$0 contains revision suggestions (Luo et al., 15 Sep 2025).

The benchmark’s annotations include target objects, start and end frames defined by a 30% contour rule, key frames, and per-frame bounding boxes. Evaluation uses mean temporal IoU and mean visual IoU, with

$$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$1

and

$$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$2

On a 2,000-instance subset, Grounded SAM 2 attains $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$3, $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$4, $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$5, and $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$6; YOLO-World attains $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$7; CG-STVG attains $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$8 and $$\text{Normalized BERTScore} = \frac{s_{model} - s_{baseline}}{s_{best} - s_{baseline}}.$$9; Grounded-VideoLLM attains $A$0 (Luo et al., 15 Sep 2025).

Dr.V-Bench covers three task formats: 3,000 yes/no QA, 6,000 multiple-choice QA, and 1,000 caption-generation QA. The benchmark is built from 15 public datasets, including ActivityNet-QA, NExT-QA, YouCook2, CLEVRER, TempCompass, Video-MME, MSVD, MSR-VTT, and VATEX (Luo et al., 15 Sep 2025). Across evaluated models, performance declines from perceptive to temporal to cognitive tasks. Representative averages are reported as 72.67% for Qwen2-VL, 67.42% for InternVL2, 77.29% for GPT-4o, 79.68% for Gemini-1.5-Pro, and 95.25% for humans (Luo et al., 15 Sep 2025).

The framework also compares favorably with Self-PEP. For GPT-4o, average performance rises from 77.29 in vanilla form to 82.33 with Self-PEP and 88.36 with Dr.V-Agent. For Qwen2-VL, the corresponding values are 72.67, 75.67, and 82.64. For LLaVA-NeXT-Video-DPO, Self-PEP is reported as harmful, decreasing 56.80 to 52.45, whereas Dr.V-Agent raises performance to 74.21 (Luo et al., 15 Sep 2025). The supplied account interprets this as evidence that external grounding and separate inconsistency reasoning can mitigate compounding internal biases more effectively than self-reflection alone.

4. Virtual Doctor as an autonomous diabetes-screening system

The “Virtual Doctor” is an interactive AI system for non-invasive prediction of type 2 diabetes mellitus and autonomous patient interaction (Spänig et al., 2019). It is implemented as a self-contained cabin that measures weight and height, records speech, conducts a standardized anamnesis, estimates T2DM probability using a deep neural network, calibrates that score into an interpretable probability, and recommends HbA1c testing when the estimated probability lies in a “twilight zone” of 30–70%.

The cabin hardware uses two 200-kg load cells with HX711 A/D converters and an Arduino Uno for weight, and an HC-SR04 ultrasonic range finder for height. With cabin height fixed at 200 cm, patient height is computed as

$A$1

where $A$2 is the head-to-ceiling distance in centimeters. BMI is then computed as

$A$3

with $A$4 in kilograms and $A$5 in meters (Spänig et al., 2019).

Speech recognition uses CMU Sphinx, specifically Pocketsphinx and Sphinxbase, with a German acoustic model, LLM, and dictionary from VoxForge. Audio is mono, 16 kHz, and 16-bit depth. The recognizer is adapted to the cabin environment and is reported to reliably handle numbers 1–100 for age, numbers 1–10 for lifestyle ratings, the categories male/female, and yes/no responses (Spänig et al., 2019).

The current prototype uses sex, age, and BMI as core predictive inputs to the DNN. It then applies post-model adjustments based on four spoken anamnesis items: polyuria, polydipsia, alcohol use on a 1–10 scale, and tobacco use on a 1–10 scale. The paper states only the directionality of these adjustments: positive polyuria or polydipsia responses increase risk, and higher alcohol or tobacco ratings increase risk. Data are stored in MySQL, and a PDF report is generated (Spänig et al., 2019).

The primary cohort is the Heinz Nixdorf Recall study with $A$6, including 2,419 female participants aged 45–75 years and 656 individuals with T2DM, corresponding to 13.6% prevalence. Known T2DM accounts for 397 participants and previously unknown T2DM for 259 participants. The primary train–test split is 80:20, feature scaling uses $A$7, and class imbalance is addressed via sub-sampling following a local case-control approach (Spänig et al., 2019).

The DNN uses 1–3 hidden layers tested, tanh hidden activations, a sigmoid output, and 100 epochs. With non-invasive inputs only—sex, age, BMI—the DNN attains $A$8, compared with $A$9 for an RBF SVM; the difference is reported as significant with $V$0 by DeLong’s method. With HbA1c added, the DNN reaches $V$1 and the SVM $V$2, with no significant difference ($V$3). Calibration with GUESS improves DNN Expected Calibration Error from 0.07 to 0.05, while SVM calibration remains approximately 0.06 before and after GUESS because kernlab already applies Platt scaling (Spänig et al., 2019).

The acceptance study appended to the system description surveyed 320 university participants after cleaning from an initial $V$4. It reports a mean intention-to-use score of 2.6 on a 1–5 scale, with men more willing than women (2.4 versus 3.0; $V$5, $V$6), younger respondents somewhat more willing than respondents aged at least 25 years (2.5 versus 2.9; $V$7, $V$8), and significant field-of-study differences ($V$9, $T$0) (Spänig et al., 2019). These results concern acceptance rather than diagnostic validity, but they show that the project combined technical development with an initial deployment-oriented social evaluation.

5. “Dr.V” in representation learning for drug response and clinical operations

A separate line of work uses “Dr.V” in model names for latent-variable and representation-learning systems. Dr.VAE, the “Drug Response Variational Autoencoder,” is a semi-supervised extension of PertVAE designed to predict treatment response from genomic profiles while learning drug-induced perturbation structure (Rampasek et al., 2017). The model learns latent gene states before and after treatment, denoted $T$1 and $T$2, and introduces a hierarchical latent variable $T$3 and binary response $T$4. Its joint distribution is

$T$5

The perturbation mapping is modeled in latent space through a residual linear Gaussian mean function,

$T$6

and the classifier is

$T$7

Inputs are 903 L1000 landmark genes, latent dimensionality is 100 for all stochastic variables, and each encoder uses two IAF steps with MADE (Rampasek et al., 2017).

Dr.VAE is evaluated on 19 drugs selected to match the AstraZeneca–Sanger DREAM Challenge panel, using GDSC and CCLE response data together with LINCS L1000 perturbation pairs. The paper reports that the VAE-based models outperform published benchmarks by 3–11% AUROC and 2–30% AUPR. Averaged across drugs, Dr.VAE achieves an 8.95% improvement over ridge logistic regression, while SSVAE achieves 8.07%. Two-step approaches such as PCA plus logistic regression or PertVAE latent features followed by logistic regression underperform the jointly trained models, and paclitaxel is identified as a notable exception where ridge logistic regression wins (Rampasek et al., 2017). The stated conclusion is that better reconstruction does not necessarily improve classification, whereas joint optimization of representation and prediction does.

The supplied account also uses “Dr.V” to denote Doctor2Vec, a doctor-vector representation-learning framework for clinical trial recruitment (Biswal et al., 2019). Doctor2Vec jointly learns trial embeddings from clinical trial text and categorical metadata and doctor embeddings from EHR-derived patient histories. Trial text is encoded with BERT, categorical trial features are encoded linearly, and the fused query is

$T$8

For each doctor, patient histories are encoded hierarchically with a visit encoder and a Bi-LSTM with attention, producing patient memory cells $T$9. Trial-conditioned attention is then

$O=\{o_1,\dots,o_n\}$0

Final prediction is given by

$O=\{o_1,\dots,o_n\}$1

Doctor2Vec is trained on 2,609 trials, 25,894 doctors, 430,239 patients, and 102,487 doctor–trial pairs. Its primary metric is PR-AUC for five-class recruitment prediction, with a reported score of $O=\{o_1,\dots,o_n\}$2, compared with $O=\{o_1,\dots,o_n\}$3 for the best baseline LSTM, corresponding to an 8.7% relative improvement. In transfer settings, the model reports a 13.7% PR-AUC gain for country transfer from the United States to South Africa and an 8.1% gain for rare diseases (Biswal et al., 2019). Although this work is not titled “Dr.V,” the supplied account explicitly groups it under that abbreviation, reinforcing that “Dr.V” can denote representation-learning systems far removed from medical vision or interactive diagnosis.

6. Recurring design patterns and limitations across usages

Despite their heterogeneity, these “Dr.V” usages share a recurrent methodological pattern: they decompose a high-level clinical or perceptual task into structured intermediate representations. DrVD-Bench organizes reasoning into levels from image quality to clinical interpretation (Zhou et al., 30 May 2025). Dr.V for video hallucination formalizes a hierarchy from perceptive to temporal to cognitive verification (Luo et al., 15 Sep 2025). The Virtual Doctor separates measurement, anamnesis, DNN scoring, calibration, and confirmatory testing policy (Spänig et al., 2019). Dr.VAE separates pre-treatment state, perturbation dynamics, and response outcome (Rampasek et al., 2017). Doctor2Vec separates visits, patients, doctors, and trials through a hierarchical memory-attention architecture (Biswal et al., 2019). This suggests that “Dr.V”-labeled systems often prioritize staged inference over monolithic prediction.

A second commonality is the reliance on proxy evidence and benchmarked intermediate signals rather than direct causal validation. DrVD-Bench measures reasoning quality by accuracy, erasure sensitivity, and normalized BERTScore, but explicitly notes that benchmark performance is not sufficient for clinical deployment (Zhou et al., 30 May 2025). Dr.V diagnoses hallucination using external grounding tools whose own performance ceilings constrain end results (Luo et al., 15 Sep 2025). The Virtual Doctor estimates diabetes probability from sex, age, and BMI, with optional HbA1c benchmarking, but does not report sensitivity, specificity, or confusion matrices (Spänig et al., 2019). Dr.VAE predicts response from cell-line perturbation and outcome data that are sparse and heterogeneous (Rampasek et al., 2017). Doctor2Vec inherits the biases of historical recruitment patterns and claims/EHR distributions (Biswal et al., 2019).

A third source of confusion is nominal rather than technical. “Dr.V” may refer to a benchmark goal, a framework name, a clinical device, or a generative model family. In the medical-image setting, there is explicitly “no separately defined ‘Dr.V’ model”; DrVD-Bench is the evaluation suite for that goal (Zhou et al., 30 May 2025). In the video setting, by contrast, Dr.V directly names both the benchmark and the agent (Luo et al., 15 Sep 2025). In older biomedical machine learning, the abbreviation instead appears inside task-specific names such as Dr.VAE (Rampasek et al., 2017). Careful citation is therefore necessary whenever the term is used in research discourse.

Taken together, the literature indicates that “Dr.V” is best understood as a context-dependent label spanning clinically inspired reasoning benchmarks, safety-oriented multimodal auditing, autonomous screening interfaces, and biomedical latent-variable models. The unifying thread is not a shared architecture but a shared aspiration toward structured, evidence-linked inference in domains where ungrounded predictions are consequential (Zhou et al., 30 May 2025, Luo et al., 15 Sep 2025, Spänig et al., 2019, Rampasek et al., 2017, Biswal et al., 2019).