KASVIR Dataset in Biomedical Imaging

• KASVIR is a dual-purpose biomedical imaging dataset featuring both urine sediment images for AI-based analysis and a lineage supporting gastrointestinal VQA benchmarks.
• The urine sediment portion comprises 8509 manually labeled particle images across eight clinically relevant classes, acquired using standardized protocols.
• The gastrointestinal usage underlies a VQA benchmark with image–question–answer triplets, facilitating multimodal clinical reasoning in endoscopic imaging.

Searching arXiv for papers mentioning KASVIR / Kvasir-VQA to ground the article. KASVIR denotes a dataset name used in two distinct biomedical-imaging contexts in the arXiv literature. In its clearest and most explicit usage, it refers to a public urine sediment image dataset created for AI-based automatic urine sediment examination (USE), containing 8509 manually labeled particle images from 409 patients and organized into 8 particle classes (Tuncer et al., 2023). In a later gastrointestinal visual question answering study, the term appears in connection with the Kvasir-VQA benchmark and the broader gastrointestinal endoscopy image repository lineage derived from the extended HyperKvasir collection, where the data take the form of image–question–answer triplets rather than particle crops (Gaihre et al., 19 Jul 2025). The term therefore spans two medically distinct but methodologically related dataset roles: structured clinical image classification in urinalysis and multimodal VQA in gastrointestinal endoscopy.

1. Terminology and referential scope

Published usage indicates that “KASVIR” is not entirely univocal. One usage names a urine sediment dataset designed for automatic particle recognition in routine urinalysis (Tuncer et al., 2023). A second usage appears in the MEDVQA-GI 2025 context, where the paper itself uses the name Kvasir-VQA for the benchmark but thanks KASVIR dataset providers in the acknowledgments; from context, KASVIR there refers to the gastrointestinal endoscopy image repository or dataset lineage underpinning the VQA benchmark (Gaihre et al., 19 Jul 2025).

Usage of the term	Clinical domain	Data unit
KASVIR urine sediment dataset	Urine sediment microscopy	Particle images
Kvasir-VQA / KASVIR lineage	Gastrointestinal endoscopy	Image–question–answer triplets

This naming overlap is a source of potential confusion. A common misconception is to treat all occurrences of “KASVIR” as referring to a single dataset. The published record instead supports a narrower interpretation: the urine sediment resource is a standalone labeled microscopy dataset, whereas the gastrointestinal usage identifies the data source lineage behind a VQA benchmark rather than an eight-class urine corpus.

2. Urine sediment dataset: clinical motivation and acquisition protocol

The urine sediment KASVIR dataset was introduced to support automatic USE, a diagnostic procedure used in the evaluation of kidney disease, urinary tract disease, metabolic disorders, and diabetes (Tuncer et al., 2023). The motivating problem is that manual microscopy is labor-intensive, time-consuming, subjective, and imprecise. The source paper further characterizes urine sediment inspection as difficult because sediment images are complex, particle appearance varies significantly, classification is operator-dependent, and the process is slow and not fully objective.

The dataset was compiled from 409 patients at Elazığ Fethi Sekin Central Hospital / Research and Training Hospital, specifically 143 women and 266 men. The laboratory workflow is described in procedural detail. Urine samples were examined within the first hour after collection. For microscopy, 10 mL of urine was placed into a centrifuge tube, centrifuged at 1500 rpm for 5 minutes to precipitate the sediment, the supernatant was poured off, and the remaining sediment was transferred to a clean slide and covered with a coverslip (Tuncer et al., 2023).

Image capture used an Optika B293PLi microscope with a trinocular brightfield setup, a high-speed, user-friendly camera, a 10 MP CMOS sensor, and USB 3.0 connection. Each analyzed image is reported as 3664 × 2748 pixels in color JPG format. No further image normalization, resizing, augmentation, or computational preprocessing is described in the available paper record. This suggests that the dataset’s primary standardization lies in sample preparation and image acquisition rather than in post hoc computational harmonization.

3. Annotation schema and particle taxonomy

The urine sediment dataset contains 8509 particle images organized into 8 classes, with each particle labeled by a specialist in biochemistry (Tuncer et al., 2023). The paper states that the “most common particles” in each sediment image were compiled and labeled. The result is a manually curated and clinically informed labeled resource rather than an automatically generated corpus.

The eight classes are defined as follows.

• Erythrocyte: red blood cells. Their presence in urine may indicate hematuria and possible bleeding or renal or urinary pathology.
• Leukocyte: white blood cells. Elevated counts can indicate inflammation or infection, especially urinary tract infection.
• Epithelial: epithelial cells. A small number may be normal, but increased numbers can indicate infection, inflammation, or malignancy.
• Bacteria: normally absent in healthy urine. Their presence suggests bacteriuria and possible urinary tract infection.
• Yeast: may appear due to contamination or true fungal infection, more often in women, and may be associated with yeast infection.
• Cylinders: also called casts. These are cylindrical proteinaceous structures formed in renal tubules and can provide clues about kidney disease.
• Crystals: urinary crystals formed from dissolved substances under certain pH or concentration conditions. Their presence and shape may be clinically relevant or normal depending on type.
• Others: a miscellaneous category including particles such as sperm, etc.

The paper states that class-wise counts are presented in Table 1, but the exact numeric values are not recoverable from the available text. What is explicit is the overall scale: 8509 total particles across 8 total classes. The presence of an “Others” category is methodologically important because it acknowledges residual visual heterogeneity that does not fit the seven more specific particle labels.

4. Methodological role in AI-based urinalysis

The urine sediment KASVIR dataset was created to provide a structured image collection for training and evaluating machine learning / deep learning systems for automatic particle detection and classification (Tuncer et al., 2023). Its stated importance lies in being a public, labeled urine sediment image resource with 8 clinically relevant particle classes, suitable for identifying particles in urine images, training machine learning models, deep learning-based classification, and educational use for biochemistry laboratory personnel.

The intended significance is operational as well as algorithmic. The authors explicitly connect the dataset to reducing dependence on manual microscopy and to improving speed, consistency, and scalability of urine sediment examination. Because the acquisition protocol is standardized and the labels are specialist-assigned, the dataset functions as a benchmarkable substrate for supervised learning. A plausible implication is that it can support experiments on inter-class morphological discrimination under realistic laboratory imaging conditions rather than under heavily curated synthetic conditions.

At the same time, the absence of further reported preprocessing details means that downstream studies must make their own choices on resizing, augmentation, normalization, and train/validation/test partitioning. This does not limit the dataset’s utility, but it does mean that methodological comparability across follow-on studies depends substantially on external protocol design rather than on a canonical preprocessing recipe encoded in the dataset description itself.

5. KASVIR in gastrointestinal VQA: the Kvasir-VQA lineage

A separate usage of the term appears in the gastrointestinal diagnostics literature, where KASVIR is associated with the endoscopy image repository lineage underlying Kvasir-VQA, the benchmark used in the ImageCLEFmed MEDVQA-GI 2025 Challenge (Gaihre et al., 19 Jul 2025). In that setting, the task is not particle classification but Visual Question Answering (VQA) over gastrointestinal endoscopic images.

According to the paper, the benchmark is derived from the extended HyperKvasir image repository and contains approximately 58,849 image–question–answer (IQA) triplets linked to about 6,500 high-resolution gastrointestinal endoscopy images. Each IQA sample consists of an image, a source label selected from six predefined categories, a question in natural language, and an answer that directly addresses the question. The questions cover diagnostic, anatomical, and procedural aspects. The dataset includes 20 unique question templates and 502 unique answers.

The paper emphasizes strong answer imbalance. Frequent answers include “none,” “no,” “yes,” and “0”, whereas clinically specific answers such as “colonoscopy” and “polyp” occur less often. Answer length is also skewed toward short responses, with most answers being single-word outputs and only a minority being multi-word. The authors explicitly interpret the benchmark as partly classification-like but still requiring generative answering for longer, more specific responses. This is a materially different data regime from the urine sediment dataset: the basic supervision unit is a multimodal language task rather than an image-level particle label.

6. Benchmark behavior, evaluation, and limitations in the VQA setting

In the MEDVQA-GI 2025 study, the Kvasir-VQA/KASVIR benchmark is used for open-ended answer generation, not fixed-class prediction (Gaihre et al., 19 Jul 2025). For computational reasons, the authors train on a 1% stratified subset. Their procedure first divides the subset into 90% training and 10% testing; the testing portion is then further split into 90% training and 10% validation, yielding the final train/validation/test configuration.

The model backbone is Florence-2, specifically microsoft/Florence-2-base-ft. The ViT-L/14 vision tower is kept frozen throughout fine-tuning. Inputs are processed with the model’s AutoProcessor, answers are tokenized with padding replaced by -100 for causal loss computation, and no parameter-efficient adapters such as LoRA are used. The reported supervised objective is

$$\mathcal{L} = - \sum_{i=1}^{|y|} \log P_\theta (y_i \mid y_{<i}, x)$$

where $\theta$ represents model parameters, $x$ is the combined input of image and question, and $y$ is the target answer sequence.

The architecture description states that the frozen vision encoder is pretrained at 896×896 resolution with 16×16 patches and produces 196 visual tokens per image, projected into a 196×1024 feature map with learned 2D positional embeddings. The decoder is a 2.7B parameter causal LLM with 32 transformer layers, 32 attention heads, and hidden size 2048, with cross-attention inserted every fourth layer. Inference uses temperature sampling with $T = 0.7$. Because the dataset lacks region annotations, location tokens for spatial grounding were not used.

Training uses AdamW with $\beta_1=0.9$, $\beta_2=0.999$, learning rate $7.8 \times 10^{-6}$, cosine decay, 20 epochs, and weight decay 0.1. The effective batch size is 20. Regularization includes gradient clipping with max norm 1.0 and dropout $p=0.1$ on attention weights. Evaluation uses BLEU, ROUGE-1, ROUGE-2, ROUGE-L, and METEOR, computed with the HuggingFace evaluate library. The paper also reports paired t-tests for baseline comparisons with $p < 0.05$.

Performance is reported as follows. On the validation set, the fine-tuned Florence model on the 1% stratified subset achieves BLEU 0.12, ROUGE-1 0.78, ROUGE-2 0.09, ROUGE-L 0.77, METEOR 0.42. On the public test set, the scores are BLEU 0.150, ROUGE-1 0.810, ROUGE-2 0.100, ROUGE-L 0.800, METEOR 0.440. On the private test set, the best reported values are BLEU 0.160, ROUGE-1 0.880, ROUGE-2 0.100, ROUGE-L 0.880, METEOR 0.490. The augmentation ablation reports no augmentation at BLEU 0.00, ROUGE-L 0.63, METEOR 0.31, heavy augmentation at ROUGE-L 0.48, METEOR 0.25, standard augmentation at BLEU 0.12, ROUGE-L 0.77, METEOR 0.42, and fine-tuned augmentation at BLEU 0.15, ROUGE-L 0.80, METEOR 0.44. The paper further notes stronger performance on “where” and “have” questions and weaker performance on “how” and “is” questions.

These results are accompanied by explicit limitations. The authors describe the benchmark as useful but challenging because it is answer-imbalanced, mostly short-answer oriented, and clinically specific. They note that overall performance remains moderate, BLEU scores are low across categories, and the model may capture semantic intent without matching the exact reference phrasing. They also highlight limited compute and data access, especially in low- and middle-income country settings, and recommend future work on visual grounding, unanswerable questions, incorporation of medical knowledge, and conversational or multi-turn extensions. In encyclopedic terms, this places the gastrointestinal usage of KASVIR within a broader shift from static medical image labeling toward multimodal clinical reasoning benchmarks, while preserving the central role of curated dataset design.