The reviewed record of science sign in
Pith

arxiv: 2606.25246 · v1 · pith:ZEDOFNJE · submitted 2026-06-24 · cs.CV · cs.CL

Multilingual Hematology Visual Question Answering Dataset

Hajra Malik , Hafiza Tooba Aftab , Abdul Rehman , Mohsen Ali , Waqas Sultani This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-25 21:06 UTCgrok-4.3pith:ZEDOFNJErecord.jsonopen to challenge →

classification cs.CV cs.CL
keywords multilingual VQAhematologyleukemiawhite blood cellsUrduvisual question answeringbilingual datasetmedical imaging
0
0 comments X

The pith

A bilingual English-Urdu dataset supplies 110,000 question-answer pairs for 20,000 white blood cell images.

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

The paper aims to address the English-only focus of existing medical vision-language resources by building a benchmark that supports visual question answering in both English and Urdu for leukemia and normal white blood cells. A survey of healthcare professionals showed frequent mismatches between English-based clinical systems and Urdu used in patient communication, particularly in South Asia. The authors construct the benchmark by combining morphology annotations from two existing cell datasets with a specialized Urdu hematology dictionary to generate consistent bilingual pairs. This creates a resource of 110K question-answer pairs across 20K single-cell images, along with baseline evaluations of open-source vision-language models.

Core claim

The authors introduce WBCMor VQA as a clinically validated bilingual benchmark containing 110K English-Urdu question-answer pairs that annotate 20K single-cell images of leukemic and normal white blood cells. The benchmark is assembled from morphology-aware annotations in prior datasets and supported by a domain-specific Urdu hematology dictionary to preserve clinical accuracy and linguistic consistency. Baseline performance results from multiple open-source vision-language models are reported on the new resource.

What carries the argument

The WBCMor VQA benchmark, built by repurposing morphology annotations from existing datasets and extending them with an Urdu hematology dictionary to produce bilingual question-answer pairs.

If this is right

  • Open-source vision-language models can be tested and trained on medical image questions that include Urdu terminology.
  • The resource supports AI systems capable of handling both English documentation and Urdu patient communication in hematology.
  • The construction method offers a pattern for generating similar bilingual benchmarks in other medical imaging domains.
  • The 20K annotated images provide a base for additional tasks such as cell classification or report generation in two languages.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Performance differences between English and Urdu questions on the benchmark could highlight where current models struggle with technical medical terms across languages.
  • If the dictionary approach scales, similar resources could be built for other languages facing English-dominant medical AI systems.
  • Integration into clinical tools might allow direct Urdu responses to image-based queries without separate translation steps.

Load-bearing premise

Repurposing annotations from prior datasets and adding a new Urdu dictionary yields clinically accurate and consistent bilingual question-answer pairs.

What would settle it

An independent review by hematologists and Urdu-speaking clinicians that finds a high rate of clinical inaccuracies or inconsistencies in a random sample of the generated question-answer pairs would show the benchmark lacks clinical validation.

Figures

Figures reproduced from arXiv: 2606.25246 by Abdul Rehman, Hafiza Tooba Aftab, Hajra Malik, Mohsen Ali, Waqas Sultani.

Figure 1
Figure 1. Figure 1: Pipeline of the proposed bilingual WBC-VQA dataset construction and model development pipeline. Cell [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Survey findings related to physician communication practices and challenges in AI-assisted healthcare systems. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prompt template used for morphology-based [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of bilingual (English and Urdu) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Vision Language Models (VLMs) have shown promising capabilities in medical image analysis by jointly understanding visual and textual information for tasks such as Visual Question Answering. However, existing hematology vision-language resources remain predominantly English centric, limiting their applicability in multilingual healthcare environments. This challenge is releveant generally to South Asia and specifically to Pakistan, where Urdu is widely used despite healthcare information and digital medical systems being largely dependent on English. To investigate this gap, we conducted a survey among healthcare professionals, which revealed substantial language mismatches between clinical documentation and patient communication, emphasizing the need for multilingual healthcare technologies. To address this limitation, we introduce WBCMor VQA, a clinically validated bilingual English, Urdu morphology aware VQA benchmark for leukemia and normal white blood cell analysis. The benchmark is constructed using morphology-aware annotations from LeukemiaAttri and WBCAtt datasets and supported by a domain specific Urdu hematology dictionary to ensure linguistic consistency and clinical correctness. The final benchmark contains 110K bilingual question answer pairs serving as VQA annotations for 20K leukemic and normal single-cell images. Furthermore, we establish baseline performance by evaluating multiple open-source VLMs on the proposed benchmark. The proposed resource aims to facilitate the development of accessible and clinically relevant AI systems for multilingual healthcare environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces WBCMor VQA, a bilingual (English-Urdu) morphology-aware VQA benchmark for leukemia and normal white blood cell analysis. It is constructed from morphology-aware annotations in the LeukemiaAttri and WBCAtt datasets, augmented by a domain-specific Urdu hematology dictionary, yielding 110K bilingual QA pairs over 20K single-cell images. The work also reports results from a survey of healthcare professionals on language mismatches and provides baseline evaluations of several open-source VLMs on the new benchmark.

Significance. If the construction process and clinical validation can be shown to be sound and reproducible, the resource would address a documented gap in multilingual medical VQA resources for Urdu-speaking regions. The bilingual design and morphology focus could support development of more accessible clinical AI tools; the reported baselines would serve as a useful reference point for subsequent work.

major comments (2)
  1. [Abstract] Abstract: the central claim that the benchmark is 'clinically validated' and ensures 'clinical correctness' rests on the repurposing of annotations from LeukemiaAttri/WBCAtt plus a new Urdu dictionary, yet no generation procedure, translation rules, expert review protocol, or quantitative validation metrics (e.g., agreement scores) are described. This information is required to assess the validity of the 110K QA pairs.
  2. [Dataset construction] Dataset construction description (wherever presented): the mapping from the 20K images and existing morphology annotations to the final 110K bilingual QA pairs is not specified. Details on question templates, how morphology awareness is preserved in Urdu, dictionary application rules, and any filtering or quality-control steps are absent, preventing evaluation of reproducibility and clinical fidelity.
minor comments (1)
  1. [Abstract] Abstract: 'releveant' is a typographical error and should read 'relevant'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments correctly identify that the current manuscript lacks explicit descriptions of the QA-pair generation process, translation methodology, expert review steps, and quantitative validation metrics. We will revise the manuscript to supply these details in a dedicated methods subsection, thereby strengthening the claims of clinical validation and reproducibility.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the benchmark is 'clinically validated' and ensures 'clinical correctness' rests on the repurposing of annotations from LeukemiaAttri/WBCAtt plus a new Urdu dictionary, yet no generation procedure, translation rules, expert review protocol, or quantitative validation metrics (e.g., agreement scores) are described. This information is required to assess the validity of the 110K QA pairs.

    Authors: We agree that the abstract and main text currently assert clinical validation without providing the supporting procedural details. The benchmark re-uses morphology annotations already present in LeukemiaAttri and WBCAtt (which were produced by hematologists) and augments them with a newly compiled Urdu hematology dictionary; however, the exact template-based question generation rules, how morphology terms were mapped while preserving clinical meaning in Urdu, the expert review protocol used to verify the dictionary, and any inter-annotator agreement statistics are not reported. We will add a new subsection (e.g., Section 3.2) that documents these steps, including the dictionary construction process and any quantitative checks performed. revision: yes

  2. Referee: [Dataset construction] Dataset construction description (wherever presented): the mapping from the 20K images and existing morphology annotations to the final 110K bilingual QA pairs is not specified. Details on question templates, how morphology awareness is preserved in Urdu, dictionary application rules, and any filtering or quality-control steps are absent, preventing evaluation of reproducibility and clinical fidelity.

    Authors: The referee is correct that the mapping procedure is underspecified. The 110K bilingual pairs are generated by applying a fixed set of morphology-aware question templates to the existing attribute annotations, followed by automatic translation via the domain dictionary and a small number of manual corrections. Because these templates, dictionary application rules, and quality-control filters are not described, reproducibility cannot be assessed from the current text. We will expand the dataset-construction section with pseudocode or explicit rules for template instantiation, the dictionary lookup procedure, and the filtering criteria applied to remove low-quality or duplicate pairs. revision: yes

Circularity Check

0 steps flagged

Dataset construction paper with no derivation or prediction chain

full rationale

This paper introduces a new VQA benchmark by repurposing annotations from LeukemiaAttri and WBCAtt plus a domain-specific Urdu dictionary. No mathematical derivations, predictions, fitted parameters, or uniqueness theorems are present. The abstract and text describe dataset assembly at a high level without any step that reduces by construction to prior inputs. No self-citation load-bearing claims or ansatz smuggling occur. This matches the default non-circular outcome for resource papers.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The contribution rests on repurposing two existing image annotation datasets and introducing a new domain dictionary; no free parameters or invented entities are introduced.

axioms (2)
  • domain assumption Morphology-aware annotations from LeukemiaAttri and WBCAtt datasets are suitable for generating VQA pairs.
    The benchmark is constructed using these annotations.
  • domain assumption A domain-specific Urdu hematology dictionary produces linguistically consistent and clinically correct translations.
    The benchmark is supported by this dictionary.

pith-pipeline@v0.9.1-grok · 5765 in / 1386 out tokens · 30594 ms · 2026-06-25T21:06:35.150778+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 8 canonical work pages · 4 internal anchors

  1. [1]

    2016 , howpublished =

    Malaria Microscopy Quality Assurance Manual , author =. 2016 , howpublished =

    2016

  2. [2]

    Scientific Data , volume=

    A dataset of clinically generated visual questions and answers about radiology images , author=. Scientific Data , volume=

  3. [3]

    Medical Image Analysis , pages=

    Leveraging sparse annotations for leukemia diagnosis on the large leukemia dataset , author=. Medical Image Analysis , pages=. 2025 , publisher=

    2025

  4. [4]

    2021 IEEE 18th international symposium on biomedical imaging (ISBI) , pages=

    Slake: A semantically-labeled knowledge-enhanced dataset for medical visual question answering , author=. 2021 IEEE 18th international symposium on biomedical imaging (ISBI) , pages=. 2021 , organization=

    2021

  5. [5]

    PathVQA: 30000+ Questions for Medical Visual Question Answering

    Pathvqa: 30000+ questions for medical visual question answering , author=. arXiv preprint arXiv:2003.10286 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  6. [6]

    MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs

    MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs , author=. arXiv preprint arXiv:1901.07042 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1901

  7. [7]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Systematic inequalities in language technology performance across the world’s languages , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  8. [8]

    AMA Reports , year =

    American Medical Association , title =. AMA Reports , year =

  9. [9]

    Journal of Medical Systems , year =

    Ali, Ahmed and Khan, Sara , title =. Journal of Medical Systems , year =

  10. [10]

    Blood Cancer Facts and Statistics , year =

  11. [11]

    JCO Clinical Cancer Informatics , volume=

    Cancer statistics in Pakistan from 1994 to 2021: data from cancer registry , author=. JCO Clinical Cancer Informatics , volume=. 2023 , publisher=

    1994

  12. [12]

    2026 , journal=

    Trends in cancer burden in Pakistan: a 30-year analysis from the Global Burden of Disease Study (1990--2019) , author=. 2026 , journal=

    1990

  13. [13]

    PSLM: Every Third Pakistani Still Illiterate , year =

  14. [14]

    Literacy Rate – 7th Pakistan Population and Housing Census , year =

  15. [15]

    , author=

    Bytes to blood: Artificial intelligence in leukemia management—A 2025 update. , author=. Journal of Clinical Oncology , volume=. 2026 , publisher=

    2025

  16. [16]

    South Asian Social Review , year=

    Linguistic Diversity and Language Access in South Asian Healthcare , author=. South Asian Social Review , year=

  17. [17]

    arXiv preprint arXiv:2511.04727 , year=

    IndicVisionBench: Benchmarking Cultural and Multilingual Understanding in VLMs , author=. arXiv preprint arXiv:2511.04727 , year=

    work page arXiv

  18. [18]

    , author=

    Ethnologue: Languages of the world. , author=. Language , volume=. 2008 , publisher=

    2008

  19. [19]

    arXiv preprint arXiv:2407.11383 , year=

    TM-PATHVQA: 90000+ textless multilingual questions for medical visual question answering , author=. arXiv preprint arXiv:2407.11383 , year=

    work page arXiv

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Wbcatt: A white blood cell dataset annotated with detailed morphological attributes , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    A large-scale multi domain leukemia dataset for the white blood cells detection with morphological attributes for explainability , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2024 , organization=

    2024

  22. [22]

    2026 , journal=

    WBC-CLIP: A Multimodal Vision-Language Framework for Morphology Aware White Blood Cell Analysis , author=. 2026 , journal=

    2026

  23. [23]

    Proceedings of CLEF (Conference and Labs of the Evaluation Forum) 2019 Working Notes , year=

    Vqa-med: Overview of the medical visual question answering task at imageclef 2019 , author=. Proceedings of CLEF (Conference and Labs of the Evaluation Forum) 2019 Working Notes , year=

    2019

  24. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Uni-Hema: Unified Model for Digital Hematopathology , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Omnimedvqa: A new large-scale comprehensive evaluation benchmark for medical lvlm , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  26. [26]

    CA: a cancer journal for clinicians , volume=

    Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries , author=. CA: a cancer journal for clinicians , volume=. 2024 , publisher=

    2022

  27. [27]

    BMJ open , volume=

    Cancer in Lahore, Pakistan, 2010--2019: an incidence study , author=. BMJ open , volume=. 2023 , publisher=

    2010

  28. [28]

    Household Integrated Economic Survey (HIES) 2024–25: Social and Economic Report , institution =

    2024

  29. [29]

    No Language Left Behind: Scaling Human-Centered Machine Translation

    No language left behind: Scaling human-centered machine translation , author=. arXiv preprint arXiv:2207.04672 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2503.01863 , year=

    Vision language models in medicine , author=. arXiv preprint arXiv:2503.01863 , year=

    work page arXiv

  31. [31]

    Frontiers in artificial intelligence , volume=

    Vision-language models for medical report generation and visual question answering: A review , author=. Frontiers in artificial intelligence , volume=. 2024 , publisher=

    2024

  32. [32]

    Frontiers in Digital Health , volume=

    Medical visual question answering with multimodal: a systematic mini review (2023--2026) , author=. Frontiers in Digital Health , volume=. 2026 , publisher=

    2023

  33. [33]

    International Conference on Image Analysis and Processing , pages=

    MORE: A Framework for Stable White Blood Cell Morphological Classification and Report Generation , author=. International Conference on Image Analysis and Processing , pages=. 2025 , organization=

    2025

  34. [34]

    arXiv preprint arXiv:2601.03915 , year=

    HemBLIP: A Vision-Language Model for Interpretable Leukemia Cell Morphology Analysis , author=. arXiv preprint arXiv:2601.03915 , year=

    work page arXiv

  35. [35]

    Findings of the association for computational linguistics: acl 2024 , pages=

    Biomistral: A collection of open-source pretrained large language models for medical domains , author=. Findings of the association for computational linguistics: acl 2024 , pages=

    2024

  36. [36]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Llava-med: Training a large language-and-vision assistant for biomedicine in one day , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    Bioengineering , volume=

    Vision--language model for visual question answering in medical imagery , author=. Bioengineering , volume=. 2023 , publisher=

    2023

  40. [40]

    2016 , eprint=

    SIFT: An Algorithm for Extracting Structural Information From Taxonomies , author=. 2016 , eprint=

    2016

  41. [41]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    LoRA in LoRA: Towards parameter-efficient architecture expansion for continual visual instruction\-tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  42. [42]

    2025 , url =

    EF English Proficiency Index 2025: A Ranking of 123 Countries and Regions by English Skills , institution =. 2025 , url =

    2025

  43. [43]

    Globalisation, Societies and Education , volume=

    The EF English Proficiency Index as an international large-scale assessment: a critical cultural political economy perspective , author=. Globalisation, Societies and Education , volume=. 2025 , publisher=

    2025

  44. [44]

    15th edn

    Ethnologue: Languages of the world. 15th edn. Ed. by Raymond G. GordonJr.. Dallas: SIL International, 2005. Pp. 1,272. ISBN 155671159X. 80 (Hb). , author=. Language , volume=. 2008 , publisher=

    2005

  45. [45]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    LoRA in LoRA: Towards parameter-efficient architecture expansion for continual visual instruction tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  46. [46]

    EF English Proficiency Index 2025: A Ranking of 123 Countries and Regions by English Skills , year =

    2025