A Multi-Center Breast FNAC Whole-Slide Cytology Dataset for AI-Assisted Patch-Wise Classification Using C1–C5 Reporting Categories

Garima Jain, Abhijeet Patil, Surabhi Jain, Sanghamitra Pati, Amit Sethi, Sandeep Mathur, Pulkit Verma, Nishi Halduniya, Jatin Kashyap, Sharat Kumar, Simmi Kharb, Sunita Singh, Sucheta Devi Khuraijam, Sushma Khuraijam, Ratan Konjengbam, Arvind Kumar, Deepali Tirkey, Saurav Banerjee, Shivani Kalhan, Rakesh Kumar Gupta, Ranjana Solanki, Deepika Hemranjani, Shashank Nath Singh, Uma Handa, Manveen Kaur, B. G. Malathi, Yogender P, Niraj Kumari, Shruti Gupta, Indu R. Nair, Vidya C, Basumitra Das, Sunil Kumar Komanapalli, Ravindra Karle, Tanaya Kulkarni, Vandana Raphael, Biswajit Dey, Vaishali Gaikwad, Nilam Adhav · 2026

Read A Multi-Center Breast FNAC Whole-Slide Cytology Dataset for AI-Assisted Patch-Wise Classification Using C1–C5 Reporting Categories by Garima Jain, Abhijeet Patil, Surabhi Jain, Sanghamitra Pati, Amit Sethi, Sandeep Mathur, Pulkit Verma, Nishi Halduniya, Jatin Kashyap, Sharat Kumar, Simmi Kharb, Sunita Singh, Sucheta Devi Khuraijam, Sushma Khuraijam, Ratan Konjengbam, Arvind Kumar, Deepali Tirkey, Saurav Banerjee, Shivani Kalhan, Rakesh Kumar Gupta, Ranjana Solanki, Deepika Hemranjani, Shashank Nath Singh, Uma Handa, Manveen Kaur, B. G. Malathi, Yogender P, Niraj Kumari, Shruti Gupta, Indu R. Nair, Vidya C, Basumitra Das, Sunil Kumar Komanapalli, Ravindra Karle, Tanaya Kulkarni, Vandana Raphael, Biswajit Dey, Vaishali Gaikwad, Nilam Adhav (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30209.txt.

Introduction

Garima Jain and a large multi-center team release a dataset for AI diagnosis of breast cytology slides. Fine needle aspiration cytology samples cells through a thin needle. Pathologists grade the samples in five categories, C1 through C5. Training AI on whole-slide images needs many expert-labeled examples. This 2026 collection gathers them from tertiary centers across India.

The prospective dataset covers three hundred twenty-one patients. It holds four hundred seventy whole-slide images in total. Two standard stains split the slides between them. Scans were made at forty times magnification in a standard format. Most slides contain annotated regions verified by experts. Those regions yield over seven thousand labeled image patches. The release adds annotation files, metadata, and inspection code. At roughly nine hundred fifty gigabytes, it is hosted on Zenodo.

Catalog record

CREATORGarima Jain, Abhijeet Patil, Surabhi Jain, Sanghamitra Pati, Amit Sethi, Sandeep Mathur, Pulkit Verma, Nishi Halduniya, Jatin Kashyap, Sharat Kumar, Simmi Kharb, Sunita Singh, Sucheta Devi Khuraijam, Sushma Khuraijam, Ratan Konjengbam, Arvind Kumar, Deepali Tirkey, Saurav Banerjee, Shivani Kalhan, Rakesh Kumar Gupta, Ranjana Solanki, Deepika Hemranjani, Shashank Nath Singh, Uma Handa, Manveen Kaur, B. G. Malathi, Yogender P, Niraj Kumari, Shruti Gupta, Indu R. Nair, Vidya C, Basumitra Das, Sunil Kumar Komanapalli, Ravindra Karle, Tanaya Kulkarni, Vandana Raphael, Biswajit Dey, Vaishali Gaikwad, Nilam Adhav
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.30209

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Multi-Center Breast FNAC Whole-Slide Cytology Dataset for AI-Assisted Patch-Wise Classification Using C1–C5 Reporting Categories Garima Jain, Abhijeet Patil, Surabhi Jain, Sanghamitra Pati, Amit Sethi, Sandeep Mathur, Pulkit Verma, Nishi Halduniya, Jatin Kashyap, Sharat Kumar, Simmi Kharb, Sunita Singh, Sucheta Devi Khuraijam, Sushma Khuraijam, Ratan Konjengbam, Arvind Kumar, Deepali Tirkey, Saurav Banerjee, Shivani Kalhan, Rakesh Kumar Gupta, Ranjana Solanki, Deepika Hemranjani, Shashank Nath Singh, Uma Handa, Manveen Kaur, B. G. Malathi, Yogender P, Niraj Kumari, Shruti Gupta, Indu R. Nair, Vidya C, Basumitra Das, Sunil Kumar Komanapalli, Ravindra Karle, Tanaya Kulkarni, Vandana Raphael, Biswajit Dey, Vaishali Gaikwad, Nilam Adhav Abstract We present a multi-center breast fine needle aspiration cytology (FNAC) dataset designed for patch-wise classification using C1–C5 reporting labels. The prospective dataset includes 321 patients and 470 whole-slide images (WSIs) collected from participating tertiary medical centers in India between May 2023 and March 2026. Slides were stained using Papanicolaou (190 WSIs) or May–Grünwald–Giemsa (280 WSIs), scanned on a Hamamatsu NanoZoomer S360 at 40 × \times magnification and 0.25 microns per pixel, and stored directly in NDPI format. Across the 470 WSIs, 446 WSIs contain annotated patch regions, yielding 7,398 PNG image patches with expert-verified C1–C5 labels. The release includes NDPI WSIs, WSI-level GeoJSON annotation files, extracted patch images, de-identified metadata, a data dictionary, a validation summary, a manifest linking WSIs to Zenodo records, and code for dataset inspection and reuse. The complete dataset is approximately 950 GB and is available through Zenodo. Keywords: breast cytology, fine needle aspiration cytology, whole slide imaging, patch-wise classification, C1–C5 reporting Background and Summary Breast cancer is one of the most common malignancies worldwide and remains a major cause of cancer-related morbidity and mortality among women [ 1 , 2 ] . Early and accurate diagnosis of breast lesions is essential for timely clinical management. In many clinical settings, especially in resource-constrained regions, fine needle aspiration cytology (FNAC) remains an important method for the initial evaluation of palpable and radiologically detected breast lesions. FNAC is minimally invasive, rapid, cost-effective, and clinically useful when performed and interpreted by trained personnel [ 3 , 4 , 5 , 6 ] . Breast FNAC is commonly used alongside clinical examination and radiological assessment as part of a triple-assessment approach [ 7 , 8 ] . Cytological interpretation requires assessment of cellularity, architectural arrangement, nuclear features, cytoplasmic characteristics, background elements, and features suggestive of benignity, atypia, suspicion, or malignancy. Interpretation can be challenging because benign proliferative lesions, atypical lesions, low-grade malignancies, and poorly preserved or scanty samples may show overlapping cytomorphological features [ 9 , 10 , 11 ] . To improve standardization, the International Academy of Cytology (IAC) Yokohama System defines five breast FNAC reporting categories: C1 insufficient or inadequate, C2 benign, C3 atypical, C4 suspicious for malignancy, and C5 malignant [ 12 , 13 ] . These categories are associated with diagnostic criteria, risk of malignancy, and suggested clinical management [ 13 , 14 , 15 , 16 ] . Digital cytology and artificial intelligence (AI)-assisted analysis have the potential to support region triaging, patch-level classification, benign-versus-malignant classification, and future slide-level decision-support workflows [ 18 , 19 , 20 , 17 ] . However, development of robust models depends on access to large, diverse, well-annotated, and technically validated datasets. In cytopathology, such datasets are difficult to construct because they require expert review, careful quality control, de-identification, and coordination across medical centers [ 20 , 17 ] . Recent multi-center cytology dataset efforts in oral cytology have highlighted the value of releasing whole-slide images, annotations, metadata, and reusable code for AI development [ 25 ] . For breast cytology, established resources such as the Wisconsin Diagnostic Breast Cancer dataset and BreakHis are based on extracted nuclear features or histopathology images rather than multi-center breast FNAC WSIs with C1–C5 patch-level cytology labels [ 23 , 24 ] . Prior breast FNAC image studies have generally used smaller cytology image datasets and binary benign-versus-malignant classification, limiting reproducible benchmarking of C1–C5 breast cytology models [ 26 , 27 ] . To address this gap, we release a multi-center breast FNAC cytology dataset for patch-wise classification using C1–C5 reporting labels. The dataset includes 321 patients and 470 digitized WSIs stained using Papanicolaou (PAP) and May–Grünwald–Giemsa (MGG) protocols. It contains 7,398 expert-verified image patches extracted from diagnostically relevant regions. The category-wise patch distribution is 33 C1 patches, 3,706 C2 patches, 478 C3 patches, 402 C4 patches, and 2,779 C5 patches. The release also includes WSI-level GeoJSON annotation files, patch images, de-identified metadata files, a data dictionary, a validation summary, and a manifest linking WSI files to the corresponding Zenodo records. By providing both WSIs and extracted patches, the dataset supports direct patch-level experiments while retaining traceability from each patch to its parent WSI and spatial location. Table 1: Dataset key statistics Attribute Key statistics Number of Patients 321 Number of WSIs 470 Stain types PAP – 190, MGG – 280 Age Range – 12 to 90 years Mean – 36 years Median – 34 years Diagnosis Non-cancerous – 209 (304) Cancerous – 112 (166) Category C1 insufficient – 4 (9) C2 benign – 194 (281) C3 atypical – 11 (14) C4 suspicious for malignancy – 11 (15) C5 malignant – 101 (151) Participating centers Pandit Bhagwat Dayal Sharma Post Graduate Institute of Medical Sciences, Rohtak – 65 (73) Regional Institute of Medical Sciences, Imphal – 60 (88) Rajendra Institute of Medical Sciences, Ranchi – 43 (72) Government Institute of Medical Sciences, Noida – 35 (53) Sawai Man Singh Medical College, Jaipur – 35 (38) Government Medical College and Hospital, Chandigarh – 28 (41) Mandya Institute of Medical Sciences, Mandya – 10 (28) AIIMS, Raebareli – 12 (12) Amrita Hospital, Kochi – 7 (10) Government Medical College, Vizianagaram (GMCV) – 6 (11) Dr. Balasaheb Vikhe Patil RMC, Loni – 6 (18) NEIGRIHMS, Meghalaya – 6 (18) Lokmanya Tilak MMC, Mumbai – 8 (8) Methods We describe the sample collection, slide digitization, quality control, WSI-level annotation, patch extraction, and patch-wise labeling process used to construct the breast FNAC cytology dataset. A prospective multi-center cohort study was conducted across participating tertiary medical centers in India. Ethical clearance was obtained from all participating centers and the Indian Council of Medical Research (ICMR). The study was carried out from May 2023 to March 2026. The dataset includes breast FNAC slides from 321 patients and 470 whole-slide images (WSIs), stained using Papanicolaou (PAP) and May–Grünwald–Giemsa (MGG) protocols. The summary of the dataset is shown in Table 1 . The dataset was designed for supervised patch-wise classification according to the five breast FNAC reporting categories C1 to C5. The overall workflow for dataset preparation is shown in Figure 1 . Sample collection and slide quality control Breast FNAC samples were collected at participating medical centers as part of routine cytopathological evaluation. FNAC was performed using the standard procedure generally followed in cytology practice. Aspirated cellular material was smeared onto clean glass slides for cytological examination. Relevant patient-level and slide-level clinical details were collected using cytopathology requisition forms. A subset of slides was fixed and stained using the PAP stain, while the remaining slides were air dried and stained using the MGG stain. The final dataset includes 190 PAP-stained WSIs and

Read the complete text