Pith. sign in

REVIEW 4 major objections 6 minor 29 references

MAARTA:Multi-Agentic Adaptive Radiology Teaching Assistant

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MAARTA claims that an adaptive team of AI agents comparing expert and student eye-gaze graphs can detect the chest-X-ray finding a student missed and explain the perceptual error, lifting GPT-4o-Mini from 25% to 75% accuracy.

desk verdict A useful system design and a new simulated dataset are undermined by a circular evaluation: the benchmark likely measures graph differencing, not perceptual-error reasoning. read the letter →

arxiv 2506.17320 v1 pith:UM3NGGCZ submitted 2025-06-18 cs.CY cs.CVcs.LG

Akash Awasthi , Brandon V. Chang , Anh M. Vu , Ngan Le , Rishi Agrawal , Zhigang Deng , Carol Wu , Hien Van Nguyen This is my paper
classification cs.CYcs.CVcs.LG
keywords multi-agentLLMsystemsradiologyeducationperceptualerrorfeedbackeye-trackinggazeanalysisthoughtgraphschestX-raydatasetadaptiveagentrecruitmentlargemultimodalmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that an adaptive team of AI agents can do what outcome-based grading cannot: look at where a radiology student's eyes went, compare that visual search to an expert's, and explain why a diagnostic finding was missed. The proposed system, MAARTA, converts expert and student gaze into structured thought graphs, measures how much of the diagnostic reasoning the student skipped, and scales the number of reasoning agents to the complexity of the miss. On a simulated dataset built from expert eye-tracking data, the paper reports that the framework beats every single-agent baseline it is tested against, with the strongest model jumping from about 25% to 75% accuracy in classifying missed diagnostic findings. A sympathetic reader would care because perceptual error is the hard part of radiology training, and current AI gives a verdict without showing the student where they looked wrong.

What carries the argument

The load-bearing mechanism is the thought graph, a directed scene graph whose nodes are fixation points with spatial coordinates and durations and whose edges are transitions between fixations; each graph is split into subgraphs, one per diagnostic finding, so expert and student attention can be aligned finding by finding. The error complexity score $C_{\text{error}} = \Delta n \cdot n_S$, where $\Delta n$ counts missed findings and $n_S$ counts the student's subgraphs, decides how many Perceptual Error Teacher (PET) agents are recruited; each PET agent uses chain-of-thought prompting on one missed subgraph to decide which of three perceptual failure modes occurred, and a consolidator LLM merges the outputs with a logical OR into a final structured JSON explanation.

What would settle it

Collect eye-tracking and dictation data from real radiology trainees reading the same chest X-rays, run MAARTA on their gaze graphs, and compare the missed findings it flags against the findings the trainees actually failed to report: if the flags do not line up, or natural gaze noise breaks the graph alignment, the simulated-error results do not transfer.

Watch

Extended reading notes

Core claim

MAARTA claims that the difference between an expert's and a student's diagnostic search can be represented as a mismatch between two thought graphs, and that reasoning over each missed subgraph separately produces personalized feedback naming not only the missed finding but the mechanism of the miss: the student never fixated the region, dwelled there too briefly, or searched adequately yet misinterpreted what they saw. The paper reports that on its simulated error dataset this mechanism detection works, with GPT-4o-Mini's accuracy rising to 75.00 and F1 to 83.00 from single-agent values of 25.46 and 64.84, and that adaptively recruiting agents by the error complexity score $C_{\text{error}} = \Delta n \cdot n_S$ beats recruiting by the raw count of errors. The paper also reports that independent Perceptual Error Teacher agents outperform communicating ones, and that the right number of agents is not a pure linear function of error complexity but also depends on model size.

Load-bearing premise

Every reported gain rests on the assumption that simulated student errors, made by deleting fixations, halving dwell times, or altering report text in expert gaze data, behave like real radiology students' perceptual errors, so the measured improvements would transfer to actual trainees.

Editorial extensions

If this is right

  • If the central claim holds, AI feedback for radiology trainees shifts from grading the diagnosis to correcting the search: a student is told which region was never fixated, dwelled on too briefly, or misread despite adequate attention, matching the three documented causes of missed chest-X-ray findings.
  • Dynamic agent recruitment by $C_{\text{error}}$ beats recruiting a fixed number of agents per error, so the system's compute budget tracks the difficulty of the case rather than a fixed overhead.
  • The reported accuracy gain holds at essentially equal response time (13.17 s vs 13.42 s for GPT-4o-Mini), so the multi-agent design adds reasoning capacity without adding latency on this benchmark.
  • Inter-agent communication degraded performance sharply (F1 12.00 vs 82.00 without communication), indicating that independent parallel PET analyses, merged only at the end, is the design that works for this task.
  • Because gaze data is turned into graph prompts, the approach does not require retraining the underlying models; it is a prompting-and-orchestration layer that can be mounted on any sufficiently strong multimodal LLM.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extending beyond the paper: if real novices' gaze noise is comparable to the simulated perturbations, eye-tracking hardware could turn any reading room into a training room, addressing the mentorship bottleneck the introduction describes without consuming more expert time.
  • The paper's own ablation shows the agent-count rule is not purely linear once model size varies, so a model-calibrated version of $f(C_{\text{error}})$ is a testable extension that might bring small models closer to the large-model gains.
  • The same expert-vs-student gaze-graph comparison could transfer to other visual search expertise tasks, such as mammography, pathology screening, or non-medical inspection work, wherever missed findings trace to where attention went.
  • The surprising result that communicating agents underperform independent ones suggests the perceptual-error task rewards diversity of analysis over consensus, a hypothesis testable in other multi-agent educational settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. MAARTA (Multi-Agentic Adaptive Radiology Teaching Assistant) is a proposed multi-agent LMM framework that converts expert and student radiology gaze/report data into structured 'thought graphs,' compares expert and student subgraphs to detect missed findings, and recruits Perceptual Error Teacher (PET) agents whose number scales with an error complexity score Cerror = Δn·nS. The authors simulate three types of perceptual errors from the EGD-CXR dataset (missed fixations, reduced fixation durations, altered transcriptions), evaluate MAARTA against single-agent baselines on a multilabel classification task, and report large gains in Table 1 (e.g., GPT-4o-Mini accuracy 75.00 vs. 25.46/23.57; F1 83.00 vs. 64.84/63.94). The paper also presents qualitative examples and ablations of adaptive agent selection and agent communication.

Significance. If the findings were established, the paper would contribute a useful direction: using eye-tracking data to give process-oriented feedback in radiology education rather than only outcome-based feedback. The authors provide a simulated dataset and an AutoGen-based implementation, and they test a reasonable set of base models. The main claim, that adaptive multi-agent reasoning detects missed diagnostic findings and explains perceptual errors, is timely. However, the current evaluation does not support the claimed mechanism because the benchmark is constructed from the same graph perturbations that the system is asked to detect, and no direct evaluation of explanation quality is provided. The contribution is therefore a promising system proposal and dataset artifact rather than a validated educational intervention.

major comments (4)
  1. [Section 2.4, Eq. (3)] The evaluation is circular with respect to the central claim. Ground-truth labels are exactly the injected perturbations: missed-fixation cases delete finding fixations, reduced-fixation cases halve dwell times, and incomplete-knowledge cases alter transcription text. MAARTA then receives both the intact expert thought graph GT and the edited student graph GS (Section 2.5), and the PET agents are instructed to compare gaze subgraphs. A trivial oracle that reports every finding subgraph present in GT but absent or with less than half the expert dwell time in GS would regenerate the labels without any LLM reasoning or adaptive agent recruitment. Consequently, the large accuracy/F1 gains in Table 1 may reflect engineered graph differencing rather than the claimed ability to 'explain how and why errors occur.' To establish the mechanism, the authors need a non-circular evaluation: real student gaze data, or at minimum held-out perturbations that are not directly recoverable from the graph encoding, plus human rating of the generated explanations.
  2. [Section 4, Table 1] The adaptive component is not formally specified. Equation (3) defines Nagents = f(Cerror), but f is never given; the text says 'a linear relationship' is assumed and agents are 'recruited directly based on Cerror,' with no definition of the exact mapping, maximum agent count, or thresholds. Table 2 compares 'Agents Based on Number of Errors' with 'Agents Based on Error Complexity Function,' but neither strategy is described in enough detail to determine what was implemented. Without a precise f and an independent validation (e.g., comparison across fixed counts, random allocations, and multiple model sizes), the claim that complexity-adaptive agent recruitment improves performance is not testable.
  3. [Section 4, 'Qualitative Results' and Figure 2] The quantitative results lack error bars, confidence intervals, significance tests, and sample sizes. Accuracy, precision, recall, and F1 are reported as single numbers, and it is not stated whether these are single runs or averages over seeds/samples. The claim of 'consistent performance gains' is therefore unsupported beyond point estimates. The authors should report the number of test instances per error type, multiple independent runs, and appropriate statistical tests (e.g., paired bootstrap or McNemar) for the differences in Table 1 and Table 2.
  4. [Section 4] The paper's central promise is personalized feedback that 'explains how and why errors occur,' but explanation quality is only supported by two anecdotal cases in Figure 2. No metric in Section 3 scores the correctness, usefulness, or pedagogical value of the generated explanations. Because the benchmark is synthetic and the explanations are produced by LLM prompting without human verification, the authors should add an evaluation with radiology trainees or expert radiologists rating the explanations for accuracy, relevance, and actionability, or at least a rubric-based scoring of explanation content.
minor comments (6)
  1. [Section 2.5 heading] The heading contains a typo: 'teaching Assiatnt' should be 'Teaching Assistant' (and 'assiatnt' should be 'assistant').
  2. [Section 4, Ablation] The text says 'Figure 2 (B) illustrates the relationship between error complexity score and Hamming loss,' but Figure 2 shows qualitative input/output examples; the referenced plot appears to be missing or mislabeled. This should be a separate figure with axes and a caption.
  3. [Section 3, Fixation-Transcription Mapping] The alignment function f that maps report timestamps to gaze fixations and the thresholds used to decide whether a fixation corresponds to a finding are not specified. These parameters directly affect the constructed thought graphs and the simulated error labels, so their values should be documented in the paper or supplementary material.
  4. [Section 3, Dataset] The authors state that the simulated error dataset is 'balanced' but do not report the number of samples per error type, the distribution of findings, or the number of unique images used. These statistics should be included to allow reproducibility.
  5. [Section 2.4] The claim that the linear assumption is 'empirically tested' is not supported by any experiment in the current text; the ablation in Table 2 compares two selection strategies but does not report the actual f used or the range of agent counts observed.
  6. [Abstract/Supplementary] The abstract and text promise an anonymous code and dataset link in the supplementary material, but no link appears in the submitted manuscript. The link should be included in the final version.

Circularity Check

2 steps flagged · score 7.0 of 10

Simulated-error benchmark is circular by construction: the ground-truth error categories are the same fixation-removal, duration-halving, and transcription-alteration operations that MAARTA detects by comparing expert and student thought graphs.

  1. self definitional [Section 3, 'Simulated Error Data' (Dataset & Experimentation)]
    "Error Synthesis, introducing three error types: (i) Missed Fixation, where finding fixations are removed; (ii) Reduced Fixation, where fixation durations are halved; and (iii) Incomplete Knowledge, where fixations remain unchanged, but transcriptions are altered to mimic misinterpretation."

    The ground-truth labels used to score MAARTA are these injected perturbations of the expert EGD-CXR thought graph. MAARTA is then given both the intact expert graph GT and the edited student graph GS, and the PET agents are instructed to determine whether the student 'failed to fixate on the abnormality, exhibited brief fixation duration, or demonstrated a gaze pattern indicative of incomplete knowledge' — exactly the three injected categories. The operation used to generate the label is the same operation the system is designed to discover, so the reported accuracy and F1 gains (e.g., GPT-4o-Mini F1 83.00 vs. 64.84) can be obtained by deterministic graph differencing rather than by LLM perceptual-error reasoning.

  2. self definitional [Sections 2.4 and 2.5, Equations (1)-(2), 'Principal LLM']
    "The number of missed diagnostic findings is: Δn = |nT − nS| (1) ... MAARTA ... computes the difference in the number of subgraphs Δn, and identifies the missed findings."

    A 'missed finding' in the simulated dataset is created by removing the corresponding fixation subgraph from the student graph. Thus the student graph GS lacks a subgraph that is present in the expert graph GT, and the model's own definition of a missed finding is precisely that absence. Consequently, Δn = |nT − nS| is not an independent inference about perceptual error; it is, by construction, the count of injected missed findings. The headline result reduces to reading off a precomputed graph difference, so the benchmark does not establish the claimed ability to explain 'how and why errors occur.'

full rationale

The quantitative evaluation is not an external benchmark; it is a synthetic dataset in which every ground-truth perceptual error is manufactured by applying a known operation to the expert thought graph: remove finding fixations, halve fixation durations, or alter transcription text. MAARTA receives both the intact expert graph GT and the edited student graph GS, and its Principal LLM computes Δn = |nT − nS| and identifies missed findings while PET agents compare gaze subgraphs. The labels and the features are therefore the same object: a trivial diff (finding subgraph present in GT but absent in GS; dwell time below half the expert duration; altered transcription) reconstructs the labels without any LLM reasoning or adaptive multi-agent coordination. The large gains in Table 1, especially GPT-4o-Mini's accuracy jump from 25.46/23.57 to 75.00 and F1 from 64.84/63.94 to 83.00, are therefore primarily evidence of reading graph differences rather than of perceptual-error explanation. No metric scores the correctness or usefulness of the generated explanations; only two anecdotal cases in Figure 2 are offered. The adaptive agent-scaling mechanism and latency measurements are not circular, and there is no load-bearing self-citation, but the central detection claim reduces by construction on this benchmark. A score of 7 reflects that the central quantitative support is compromised, while the LLM's ability to read the provided graphs and produce explanation text is not entirely predetermined.

Assumptions & free parameters 2 free parameters · 4 assumptions · 3 invented entities

The central claim rests on a synthetic dataset whose error categories are assumed, a graph representation whose cognitive validity is unvalidated, and an error-complexity formula that is asserted ad hoc. The two free parameters (agent scaling function and alignment thresholds) are not reported, which weakens the quantitative conclusions.

free parameters (2)
  • Agent scaling function f(Cerror) = not specified; assumed linear
    Section 2.4 states a linear relationship is assumed and agents are recruited 'directly based on Cerror', but no slope or offset is given, so the exact mapping is underdetermined and could have been tuned to fit the results.
  • Fixation-transcription alignment thresholds = not reported
    Section 2.3 and Section 3 introduce a transformation f and 'Fixation-Transcription Mapping' aligning report sentences with gaze fixations; no parameters or validation for this alignment are given, yet the thought graphs depend on it.
assumptions (4)
  • domain assumption The three perceptual error types (missed fixation, reduced fixation, incomplete knowledge) are the primary causes of missed findings in radiology.
    Section 1 cites Gefter et al. [7] for these categories and the entire simulated dataset is built on them; if real student errors are more diverse, the benchmark is unrepresentative.
  • domain assumption A thought graph, with fixation nodes and transition edges, faithfully represents diagnostic reasoning, and subgraphs correspond to discrete findings.
    Section 2.3 defines thought graphs as directed scene graphs; this mapping from gaze to structured reasoning is asserted without validation.
  • ad hoc to paper Error complexity is adequately captured by Cerror = Δn · nS.
    Equation (2) in Section 2.4 introduces this product without derivation; it is a design choice, not a derived quantity.
  • domain assumption The transformation f aligning report transcriptions with gaze fixations is accurate.
    Section 3's Fixation-Transcription Mapping yields 1,025 mapped samples, but no accuracy or human validation of the alignment is reported.
invented entities (3)
  • Thought graph
    purpose: A graph representation of gaze fixations and transitions used to prompt LLMs
    Introduced in Section 2.3; no external validation that this representation improves educational feedback.
  • PET (Perceptual Error Teacher) agent
    purpose: A specialized agent that compares student and expert gaze subgraphs to explain missed findings
    Introduced in Section 2.5; its output quality is only measured indirectly through classification accuracy, not through actual teaching effectiveness.
  • Error complexity score Cerror
    purpose: A scalar used to determine the number of agents
    Equation (2) in Section 2.4; it is an ad hoc quantity with no external benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAARTA:Multi-Agentic Adaptive Radiology Teaching Assistant." pith.science (2026). https://pith.science/paper/UM3NGGCZ

@misc{pith2026250617320,
  author       = {Pith},
  title        = {Pith review of: MAARTA:Multi-Agentic Adaptive Radiology Teaching Assistant},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UM3NGGCZ}},
  note         = {Machine review of arXiv:2506.17320}
}
read the original abstract

Radiology students often struggle to develop perceptual expertise due to limited expert mentorship time, leading to errors in visual search and diagnostic interpretation. These perceptual errors, such as missed fixations, short dwell times, or misinterpretations, are not adequately addressed by current AI systems, which focus on diagnostic accuracy but fail to explain how and why errors occur. To address this gap, we introduce MAARTA (Multi-Agentic Adaptive Radiology Teaching Assistant), a multi-agent framework that analyzes gaze patterns and radiology reports to provide personalized feedback. Unlike single-agent models, MAARTA dynamically selects agents based on error complexity, enabling adaptive and efficient reasoning. By comparing expert and student gaze behavior through structured graphs, the system identifies missed findings and assigns Perceptual Error Teacher agents to analyze discrepancies. MAARTA then uses step-by-step prompting to help students understand their errors and improve diagnostic reasoning, advancing AI-driven radiology education.

Figures

Figures reproduced from arXiv: 2506.17320 by the authors.

Figure 1
Figure 1. Overview of the proposed methodology. (A) Thought graph generation process from eye gaze data. (B) Structure of the thought graphs representing expert and student gaze patterns. (C) Key modules of MAARTA for adaptive agent recruitment. (D) Workflow of PET agents in analyzing gaze sub-patterns and identifying perceptual errors. 2 Methodology 2.1 Background: Multi-agent LLMs and LMMs represent a shift from single-mode… view at source ↗
Figure 2
Figure 2. Qualitative results: MAARTA’s reasoning in detecting missed findings [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [1]

    Accessed: 2024-11-01

    Together ai, 2022. Accessed: 2024-11-01

    work page 2022

  2. [2]

    Mandating limits on workload, duty, and speed in radiology.Radiology, 304(2):274–282, 2022

    Robert Alexander, Stephen Waite, Michael A Bruno, Elizabeth A Krupinski, Leonard Berlin, Stephen Macknik, and Susana Martinez-Conde. Mandating limits on workload, duty, and speed in radiology.Radiology, 304(2):274–282, 2022

    work page 2022

  3. [3]

    Efficient prompting methods for large language models: A survey

    K Chang, S Xu, C Wang, Y Luo, T Xiao, and J Zhu. Efficient prompting methods for large language models: A survey. arxiv 2024.arXiv preprint arXiv:2404.01077

    arXiv 2024

  4. [4]

    Are more llm calls all you need? towards the scaling properties of compound ai systems

    Lingjiao Chen, Jared Quincy Davis, Boris Hanin, Peter Bailis, Ion Stoica, Matei A Zaharia, and James Y Zou. Are more llm calls all you need? towards the scaling properties of compound ai systems. Advances in Neural Information Processing Systems, 37:45767–45790, 2025

    work page 2025

  5. [5]

    Radiology and the medical student: do increased hours of teaching translate to more radiologists?BJR| Open, 3(1):20210074, 2021

    Cindy Chew, Patrick J O’Dwyer, and David Young. Radiology and the medical student: do increased hours of teaching translate to more radiologists?BJR| Open, 3(1):20210074, 2021

    work page 2021

  6. [6]

    The llama 3 herd of models, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, and Amy Yang et al. The llama 3 herd of models, 2024

    work page 2024

  7. [7]

    Commonlymissedfindings on chest radiographs: causes and consequences.Chest, 163(3):650–661, 2023

    WarrenBGefter,BenjaminAPost,andHirotoHatabu. Commonlymissedfindings on chest radiographs: causes and consequences.Chest, 163(3):650–661, 2023

    work page 2023

  8. [8]

    Sciagents: Automating scientific dis- covery through bioinspired multi-agent intelligent graph reasoning.Advanced Ma- terials, page 2413523, 2024

    Alireza Ghafarollahi and Markus J Buehler. Sciagents: Automating scientific dis- covery through bioinspired multi-agent intelligent graph reasoning.Advanced Ma- terials, page 2413523, 2024

    work page 2024

Show all 29 references
  1. [9]

    Ghezloo et al

    F. Ghezloo et al. Pathfinder: A multi-modal multi-agent system for medical diag- nosticdecision-makingappliedtohistopathology. arXiv preprint arXiv:2502.08916, 2025

    2025 arXiv

  2. [10]

    Perception’s crucial role in radiology education

    Richard B Gunderman and Parth Patel. Perception’s crucial role in radiology education. Academic radiology, 26(1):141–143, 2019

    2019

  3. [11]

    Guo et al

    T. Guo et al. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680 , 2024

    2024 arXiv

  4. [12]

    Llm multi-agent systems: Challenges and open problems

    Shanshan Han, Qifan Zhang, Yuhang Yao, Weizhao Jin, Zhaozhuo Xu, and Chaoyang He. Llm multi-agent systems: Challenges and open problems. arXiv preprint arXiv:2402.03578, 2024

    2024 arXiv

  5. [13]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Deven- dra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Tho...

    2023

  6. [14]

    Artificial intelligence in radiology: decision support systems

    Charles E Kahn Jr. Artificial intelligence in radiology: decision support systems. Radiographics, 14(4):849–861, 1994

    1994

  7. [15]

    Krupinski, and Mehdi Moradi

    Alexandros Karargyris, Satyananda Kashyap, Ismini Lourentzou, Joy Wu, Arjun Sharma, Matthew Tong, Shafiq Abedin, David Beymer, Vandana Mukherjee, Eliz- abeth A. Krupinski, and Mehdi Moradi. Creation and validation of a chest x-ray dataset with eye-tracking and report dictation...

    2009 arXiv

  8. [16]

    Mdagents: An adaptive collaboration of llms in medical decision making

    Yubin Kim, Chanwoo Park, Hyewon Jeong, Yik Siu Chan, Xuhai Xu, Daniel Mc- Duff, Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, and Hae Won Park. Mdagents: An adaptive collaboration of llms in medical decision making. 2024

    2024

  9. [17]

    Visual search patterns and experience with radiological images

    Harold L Kundel and Paul S La Follette Jr. Visual search patterns and experience with radiological images. Radiology, 103(3):523–528, 1972. 10 Awasthi et al

    1972

  10. [18]

    Can large language models analyze graphs like professionals? a benchmark, datasets and models.Advances in Neural Information Processing Systems, 37:141045–141070, 2025

    Xin Li, Weize Chen, Qizhi Chu, Haopeng Li, Zhaojun Sun, Ran Li, Chen Qian, Yiwei Wei, Chuan Shi, Zhiyuan Liu, et al. Can large language models analyze graphs like professionals? a benchmark, datasets and models.Advances in Neural Information Processing Systems, 37:141045–141070, 2025

    2025

  11. [19]

    A survey on llm-based multi- agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on llm-based multi- agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    2024

  12. [20]

    Gpt-4 technical report, 2024

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, and Janko Altenschmidt et al. Gpt-4 technical report, 2024

    2024

  13. [21]

    Artificial intelligence-powered clinical decision support and simulation platform for radiology trainee education

    ChintanShah,KarapetDavtyan,IlyaNasrallah,RNickBryan,andSuyashMohan. Artificial intelligence-powered clinical decision support and simulation platform for radiology trainee education. Journal of Digital Imaging , 36(1):11–16, 2023

    2023

  14. [22]

    Auto- mated radiology report generation: A review of recent advances.IEEE Reviews in Biomedical Engineering, 2024

    Phillip Sloan, Philip Clatworthy, Edwin Simpson, and Majid Mirmehdi. Auto- mated radiology report generation: A review of recent advances.IEEE Reviews in Biomedical Engineering, 2024

    2024

  15. [23]

    Agentic llm workflows for generating patient-friendly medical reports

    Malavikha Sudarshan, Sophie Shih, Estella Yee, Alina Yang, John Zou, Cathy Chen, Quan Zhou, Leon Chen, Chinmay Singhal, and George Shih. Agentic llm workflows for generating patient-friendly medical reports. arXiv preprint arXiv:2408.01112, 2024

    2024 arXiv

  16. [24]

    Investi- gating the link between radiologists’ gaze, diagnostic decision, and image content

    Georgia Tourassi, Sophie Voisin, Vincent Paquit, and Elizabeth Krupinski. Investi- gating the link between radiologists’ gaze, diagnostic decision, and image content. Journal of the American Medical Informatics Association , 20(6):1067–1075, 2013

    2013

  17. [25]

    How visual search relates to visual diagnostic performance: a narrative systematic review of eye-tracking research in radiology

    A Van der Gijp, CJ Ravesloot, H Jarodzka, MF Van der Schaaf, IC Van der Schaaf, Jan PJ van Schaik, and Th J Ten Cate. How visual search relates to visual diagnostic performance: a narrative systematic review of eye-tracking research in radiology. Advances in Health Sciences Ed...

    2017

  18. [26]

    A review of perceptual expertise in radiology-how it develops, how we can test it, and why humans still matter in the era of artificial intelligence

    Stephen Waite, Zerwa Farooq, Arkadij Grigorian, Christopher Sistrom, Srinivas Kolla,AnthonyMancuso,SusanaMartinez-Conde,RobertGAlexander,AlanKan- tor, and Stephen L Macknik. A review of perceptual expertise in radiology-how it develops, how we can test it, and why humans still...

    2020

  19. [27]

    Analysis of per- ceptual expertise in radiology–current knowledge and a new perspective.Frontiers in human neuroscience, 13:213, 2019

    Stephen Waite, Arkadij Grigorian, Robert G Alexander, Stephen L Macknik, Marisa Carrasco, David J Heeger, and Susana Martinez-Conde. Analysis of per- ceptual expertise in radiology–current knowledge and a new perspective.Frontiers in human neuroscience, 13:213, 2019

    2019

  20. [28]

    Autogen: En- abling next-gen llm applications via multi-agent conversation

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: En- abling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155, 2023

    2023 arXiv

  21. [29]

    Chain of agents: Large language models collaborating on long-context tasks

    Yusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister, Rui Zhang, and Sercan Arik. Chain of agents: Large language models collaborating on long-context tasks. Advances in Neural Information Processing Systems , 37:132208–132237, 2025

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.