# A Nasal Cytology Dataset for Object Detection and Deep Learning

**Authors:** Mauro Camporeale, Giovanni Dimauro, Matteo Gelardi, Giorgia Iacobellis, Mattia Sebastiano Ladisa, Sergio Latrofa, Nunzia Lomonte  
**arXiv:** [2404.13745](https://arxiv.org/abs/2404.13745)  
**Date:** 2024-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.13745v1/

## Key points

- Nasal Cytology is a valuable clinical technique for diagnosing rhinitis and allergies
- Integration of artificial intelligence (AI) in cell counting could revolutionize the dissemination of this technique
- Introduction of the Nasal Cytology Dataset (NCD) for training Object Detection models to assist physicians and biologists
- Challenge of imbalanced distribution of cell types addressed using novel machine learning-based approach with DETR and YOLO models
- Creation process involved capturing 700 images from microscope fields, expert annotations using Roboflow platform, and stringent quality control policy implemented

## Summary

, , , , 

Nasal Cytology is a valuable clinical technique for diagnosing rhinitis and allergies. However, its widespread adoption has been hindered by the time-consuming process of manually counting cells. The integration of artificial intelligence (AI) in cell counting could revolutionize the dissemination of this technique. This article introduces the Nasal Cytology Dataset (NCD), which comprises field images of nasal cytology aimed at training Object Detection models to assist physicians and biologists in clinical practice. By replicating the real distribution of cytotypes found in the nasal mucosa through sampling images from clinical patient slides and manually annotating each cell, this dataset provides a comprehensive resource for developing AI solutions in rhinology. One significant challenge in object detection tasks related to nasal cytology is the imbalanced distribution of cell types, particularly rare ones. To address this issue, the authors present a novel machine learning-based approach using DETR and YOLO models. These models demonstrate promising performance in detecting and classifying nasal mucosa cells accurately, showcasing their potential to enhance the efficiency of rhinology experts. The dataset creation process involved capturing 700 images with a size of 1024×768 from microscope fields using a 3MP DCE-PW300 camera as per rhino-cytology protocol standards. Expert annotations were meticulously carried out using the Roboflow platform, ensuring precise labeling of each cell in every image. Additionally, a stringent quality control policy was implemented to exclude images with sampling noise, duplication issues, or overly dense clusters of cells. Overall, this study contributes significantly to addressing open challenges in automated detection and classification of nasal mucosa cells by leveraging advanced AI technologies. By facilitating faster and more accurate analysis of nasal cytology samples, these developments hold great promise for improving diagnostic processes and advancing research on respiratory conditions.