Segmentation models and experimental results on segmenting slide scans of diatom preparations from river Menne

This archive contains the models and the results of the deep learning experiments published in Kloster et al. 2022: Improving deep learning-based segmentation of diatoms in gigapixel-sized virtual slides by object-based tile positioning and object integrity constraint. The folders contain models and results of each of the 24 training runs, the files “metrics.experiments.pt[prediction score threshold].csv contain segmentation score thresholds the models obtained on unknown evaluation data. The data pertaining to each model is stored in a separate folder. Its name follows the convention “experiment.[model architecture].[tiling method].[dataset size].[timestamp]”. Please note that the naming of the tiling method differs from the manuscript; “fixed” refers to fixed-stride tiling, “objectcentred” object-based positioning, and “objectcentred_with_cropped” to object-based positioning + object integrity constraint. Dataset size “10p” refers to a 10% subsample of the complete training data set, “25p” to a 25% subsample and so on. Each folder contains the best performing model of the corresponding training run as pth (Mask R-CNN, PyTorch) or h5 (U-Net, Tensorflow/Keras) file, along with a files describing setup and conduction of the training. Subfolders “test_images_segmented*” contain the segmentation predicted by the models on the evaluation data. These are supplied as 1.) mask image either with the intensity value representing the prediction score (score 0.0 – 1.0 = intensities 0 – 255) or thresholded by the prediction score threshold given in the folder name; 2.) input image overlayed with ground truth (red) and segmentation mask (green), resulting in TP marked in yellow; 3.) a CSV file giving info on the filenames and the segmentation performance metrics. Please refer to the manuscript for further details