Biomedical Images Dataset

This repository contains the data from the manuscript "MaZda versus PyRadiomics – Comparison of the Effectiveness of Texture Feature Estimation in the Context of Classifying Selected Biomedical Images". During data collection, repositories available on the Kaggle and ISIC platforms were used, as well as the authors’ own resources prepared for earlier research. Five datasets containing images recorded using various imaging techniques were prepared. The characteristics of these datasets, including their division into observation classes, were as follows:• chest X-ray images – 2000 images divided into 4 classes (Normal, Covid-19, Viral Pneumonia, Lung Opacity);• lung CT images – 1000 images divided into 2 classes (Non-Covid, Covid-19);• brain MRI images – 2000 images divided into 4 classes (Normal, Glioma Tumor, Meningioma Tumor, Pituitary Tumor);• thyroid ultrasound images – 800 images divided into 2 classes (Normal, Hashimoto’s Disease);• dermoscopic images of skin lesions – 1600 images divided into 4 classes (Nevus, Seborrheic Keratosis, Malignant Melanoma, Basal Cell Carcinoma).

As part of the image pre-processing, cropping and conversion to an 8-bit grayscale format were performed. All datasets were randomly divided into training and testing parts (70% and 30% of the full set respectively). In all datasets, the observation classes are balanced (each class contains the same number of samples). The collection of images contained in this repository can be used for testing and comparing methods of texture feature estimation, such as measures based on the gray-level co-occurrence matrix (GLCM), LBP descriptors, or wavelet transform parameters. In addition, they provide a suitable environment for validating image classification algorithms, both traditional ones (K-NN, SVM, decision trees) and those based on deep neural networks. Thanks to the controlled experimental conditions and known class labels, it is possible to objectively compare the effectiveness of individual approaches and optimize model architectures.

Using this dataset please cite:Omiotek, Z., Boyko, O. (2026). MaZda versus PyRadiomics – comparison of the effectiveness of texture feature estimation in the context of classifying selected biomedical images. Przegląd Elektrotechniczny, 102(5), 117-126. DOI: 10.15199/48.2026.05.15.

本仓库包含论文《MaZda与PyRadiomics对比：特定生物医学图像分类场景下纹理特征提取有效性的比较》的配套数据集。在数据集采集过程中，使用了Kaggle与ISIC平台公开的数据集，以及作者团队为前期研究制备的自有资源。本次研究共制备5组数据集，涵盖不同成像技术采集的医学图像。各数据集的属性（包括观测类别划分）如下：
• 胸部X线图像：共2000张，分为4个类别（正常、新冠病毒感染、病毒性肺炎、肺部浸润）；
• 肺部CT图像：共1000张，分为2个类别（非新冠、新冠病毒感染）；
• 脑部MRI图像：共2000张，分为4个类别（正常、胶质瘤、脑膜瘤、垂体瘤）；
• 甲状腺超声图像：共800张，分为2个类别（正常、桥本甲状腺炎）；
• 皮肤皮损皮肤镜图像：共1600张，分为4个类别（痣、脂溢性角化病、恶性黑素瘤、基底细胞癌）。

图像预处理环节包含裁剪操作与8位灰度格式转换。所有数据集均按7:3的比例随机划分为训练集与测试集（分别占全量数据集的70%与30%）。所有数据集的观测类别均保持平衡（每个类别包含相同数量的样本）。

本仓库收录的图像集可用于测试与对比纹理特征提取方法，例如基于灰度共生矩阵（gray-level co-occurrence matrix, GLCM）、局部二值模式（LBP）描述符以及小波变换参数的各类方法。此外，该数据集可用于验证图像分类算法，既包括传统机器学习算法（K近邻、支持向量机、决策树），也涵盖基于深度神经网络的算法。得益于可控的实验环境与已知的类别标签，可客观对比不同方法的有效性，并优化模型架构。

若使用本数据集，请引用如下文献：Omiotek, Z., Boyko, O. (2026). MaZda versus PyRadiomics – comparison of the effectiveness of texture feature estimation in the context of classifying selected biomedical images. Przegląd Elektrotechniczny, 102(5), 117-126. DOI: 10.15199/48.2026.05.15.