Fig. 1

Schematic of data flow during whole-slide image tile extraction and image processing. a Schematic of initial slide processing and grid preparation. Whole-slide images can be annotated with Regions of Interest (ROI) to include only relevant areas of a slide for subsequent analysis. Optional slide filtering steps, including Gaussian blur filtering and Otsu’s thresholding, may be applied at this step. The remaining areas of the slide are sectioned into a grid in preparation for tile extraction. b Data flow during tile extraction. If tile-level filtering is to be performed, such as whitespace or grayspace filtering, a low-magnification image at the highest pyramidal layer is taken at the given location and used for background filtering. If the tile passes filtering, the full-magnification image extracted, optionally resized to match a target micron size, stain normalized, and converted into a PNG image, JPEG image, or Numpy array. Extracted tiles can be saved to disk as individual files, or buffered into TFRecords for faster dataset reading. c Schematic of data flow when reading from TFRecords. Image tiles are buffered in TFRecords with JPEG or PNG compression and stored with slide and location metadata. During dataset iteration, data is decoded and converted to Tensors. Augmentation, including random flipping/rotation, random JPEG compression, and random Gaussian blur, can be applied at this step. If stain normalization was not performed during tile extraction, stain normalization can be applied at this step when iterating through a TFRecord dataset in real-time. Images are then standardized, and slide names are converted to ground-truth outcome labels using a provided CSV annotations file or Pandas DataFrame