BloodyWell

Le résumé fourni par la source

The BloodyWell dataset contains 3139 anonymized images of blood samples in wells where an agglutination reaction had a potential to occur. Each image is supplied with a true label based on patient's medical history, as well as an assessment from six medicinal experts on how certain they are that the agglutination is present: certainly no agglutination (0), likely no agglutination (1), uncertain (2), likely agglutination present (3), certainly agglutination present (4). File "experts_info.json" contains data annotations, where "gt_result" is a true label, "expert_votes_positions" list contains expert annotations, and other field provide additional information. The data is split in proportions of 1:4 for training and testing machine learning models respectively, stored in files "train_dataset_10_10_80.json" and "test_dataset_10_10_80.json" respectively. The mobile_data.zip file provides out-of-distribution data for the main imgs.zip dataset. It contains 1082 images of agglutination reactions in wells, captured with multiple mobile phones. The OOD data contains repeats of the same wells under varying conditions (camera device, angle, lighting, and preprocessing pipelines). The plate types (material, well number and shape) differ from the main dataset. The corresponding labels are stored in mobile_markup.json.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.