jaharkes
commented
4 years ago I assume you are hitting an overflow somewhere, that particular slide has 4 levels internally.
- level 0: 79680 x 53443 (downsample 1)
- level 1: 19920 x 13360 (downsample 4.000112275...)
- level 2: 4980 x 3340 (downsample 16.0004491...)
- level 3: 2490 x 1670 (downsample 32.0008982...)
So it works when you are scaling starting from level 3 where you read out a 2490x1670 sized region (16MB) and scaling it down to where you want to be. But because of the 3 extra pixels on the height of level 0, the optimal level for a 32 scale factor ends up to be level 2. So then you are using read_region over a 4980x3340 image (65MB). I assume either Cairo (used to stitch the slide's internal tiles inside read_region) or PIL isn't handling the single large region very well.
I assume that if you ignore the recommendation from best_level_for_downsample and just force level = 3, it will 'upsample' it to 53443/32. = 1670.09375 which rounds down to the level 3 dimensions that were read and resize becomes a noop.
CasiaFan
Context
Issue type (bug report or feature request): Reading and resizing Openslide to PIL image creates black boxes
Details
My issue is very similar to that which someone else posted here: https://stackoverflow.com/questions/63190495/reading-and-resizing-svs-slide-results-in-a-damaged-image When I use scale factors >= 32, the image comes out fine. However, if use anything less, I get these weird black boxes.
The code to reproduce it will be from the link above. I think for this person, the scaling factor was good for 64, but didn't work for 32 or lower. Slide he / she used was here: https://portal.gdc.cancer.gov/files/5b54f176-3727-4a02-aa7a-921eebe5aeea.
def slide_to_scaled_pil_image(slide, SCALE_FACTOR=32): """ Convert a WSI training slide to a scaled-down PIL image. Args: slide: An OpenSlide object. Returns: Tuple consisting of scaled-down PIL image, original width, original height, new width, and new height. """
large_w, large_h = slide.dimensions new_w = math.floor(large_w / SCALE_FACTOR) new_h = math.floor(large_h / SCALE_FACTOR) level = slide.get_best_level_for_downsample(SCALE_FACTOR) whole_slide_image = slide.read_region((0, 0), level, slide.level_dimensions[level]) whole_slide_image = whole_slide_image.convert("RGB") img = whole_slide_image.resize((new_w, new_h), PIL.Image.BILINEAR) return img, large_w, large_h, new_w, new_h
img = openslide.OpenSlide(PATH_TO_SLIDE) ds_img, large_w, large_h, new_w, new_h = slide_to_scaled_pil_image(img, SCALE_FACTOR=SCALE_FACTOR)