Heterogeneous Medical-Image Dataset for Multi-Task ResNet-18 Analysis
Description
This dataset contains 56,900 medical-image records assembled by combining data from multiple publicly available source datasets into a unified research collection for heterogeneous medical-image analysis. The dataset covers five medical-image categories: Brain Tumor (253 images), Chest X-ray (5,856 images), Fracture (40,005 images), Heart ECG (756 images), and Heart Echo (10,030 images). The dataset record includes Final_Data.csv, which provides the image filepath, encoded primary label, data type, and source medical-image category. It also includes class_distribution.csv containing the distribution of primary labels and source categories, train_validation_split.csv containing the image-level train/validation assignments, and medical_images.zip containing the corresponding images organized by category. The associated experiment uses an 80% training and 20% validation split with stratification by the primary label and random seed 32, resulting in 45,520 training records and 11,380 validation records. The split is performed at the image-record level; patient-level or study-level grouping is not demonstrated in the experimental implementation. The Heart Echo source data originally contained echocardiography videos. For the experimental dataset, one representative middle frame was extracted from each video using Python and OpenCV. The selected frame was saved as a JPEG image and used as the corresponding image record. The resulting Heart Echo images are included in the image collection. The dataset was prepared to support research in medical imaging, computer vision, machine learning, transfer learning, and heterogeneous medical-image analysis. The original source datasets may have their own licenses, copyright conditions, attribution requirements, and redistribution restrictions. Users are responsible for reviewing and complying with the terms of the original source datasets. This dataset is intended for research and educational purposes and is not intended for direct clinical diagnosis or medical decision-making.
Files
Institutions
- Akal UniversityPunjab, Bathinda