isvisible
About
Summary
This algorithm was developed with the objective of detecting and segmenting intestinal spirochetosis (IS) in H&E-stained colon whole-slide images (WSIs). It outputs a pixel-level binary segmentation mask that identifies tissue regions potentially indicating where spirochetosis could be expected.
Mechanism
The algorithm runs automatically on a single WSI and outputs a full-slide segmentation mask at x20 magnification (0.5 µm/px).
Model architecture: DeepLabV3Plus with a Mix Transformer encoder (mit_b0), implemented via segmentation_models_pytorch.
Input: - Datatype: H&E-stained colorectal biopsy or resection whole-slide image - File formats: TIFF (OpenSlide-readable formats) - Required metadata: pixel spacing (mpp) stored in slide properties
Output:
- A full-slide binary segmentation mask (compressed TIFF) at x20 resolution
- Pixel value 255 = spirochetosis present; 0 = absent
- Delivered to /output/images/spirochetosis-mask/<slide-name>.tif
Pipeline steps:
1. Reads pixel spacing (mpp) from OpenSlide metadata to locate the x20 pyramid level.
2. Builds a tissue mask at 20 µm/px using HSV thresholding (H&E bounds: lower [90, 8, 103], upper [180, 255, 255]; minimum component area 1000 px).
3. Iterates non-overlapping 1024 × 1024 px tiles over tissue regions.
4. Normalises each tile with ImageNet mean/std and performs a forward pass through DeepLabV3Plus-mit_b0.
5. Stitches per-tile sigmoid probability maps into a full-slide float mask.
6. Thresholds at 0.5 and saves as a compressed TIFF.
Training data: The Human Intestinal Spirochetosis (HIS) dataset comprises 289 H&E-stained colonic biopsy WSIs from 78 patients with intestinal spirochetosis, sourced from the Radboudumc spirochetosis archive. Of these, 170 slides were manually annotated with expert-drawn polygons. Patches of 1024 × 1024 px were extracted at x20 resolution (0.5 µm/px) using an annotation-focused strategy. After discarding patches with no foreground annotation, 33,965 image–mask pairs were retained. The dataset was split by patient: 54 training patients (122 slides, 26,732 patches), 12 validation patients (31 slides, 5,436 patches), and 12 test patients (17 slides, 1,797 patches).
Interfaces
This algorithm implements all of the following input-output combinations:
Validation and Performance
Patch-level segmentation metrics (mean ± std) evaluated on the HIS dataset. The test set was withheld from all model development.
| Split | Dice | Precision | Recall |
|---|---|---|---|
| Validation | 0.726 ± 0.188 | 0.734 ± 0.212 | 0.806 ± 0.197 |
| Test | 0.751 ± 0.161 | 0.842 ± 0.171 | 0.717 ± 0.165 |
Ground truth: Manual pixel-level annotations (expert-drawn polygons) on H&E-stained colonic biopsy WSIs from the Radboudumc spirochetosis archive.
Uses and Directions
This algorithm is intended for research use only.
- General use: Automated screening and quantification of intestinal spirochetosis extent in H&E WSIs.
- Intended users: Researchers and computational pathology teams.
- Potential benefits: Reduces manual annotation burden; enables large-scale retrospective studies.
Warnings
- Not validated for clinical use. Do not use this algorithm to support clinical diagnosis or treatment decisions.
- The model was trained on a specific scanner/staining protocol. Performance may degrade on slides from different scanners, staining protocols, or tissue preparation methods not represented in the training set.
- Slides missing pixel spacing metadata (mpp) in their OpenSlide properties will fail with an error. Ensure scanner metadata is present before submission.
- Very large slides (> several gigapixels) may exceed the memory or time limits of the Grand Challenge runtime instance.