Limitations
What this tool can and cannot tell you. Read this before you rely on a result.
This result is an aid, not a diagnosis. Segmentation groups cells, it does not isolate individual cells.
- Segmentation groups cells, it does not isolate individual cells. The object count is therefore a rough number, and cell shape from segmentation is not validated yet.
- The test result is from 66 images of the same public dataset as the training images. It is a sanity check, not clinical validation, and says nothing about new strains.
- Only two shapes are predicted, cocci and bacilli. DIBaS has no spiral species, so a spiral class is not predicted.
- The model learned from Gram stained images like those in DIBaS. Images from other stains, microscopes or organisms can give unreliable answers, so do not trust the confidence level alone on them.
- Use PNG, JPG or TIFF up to 20 MB, with one in-focus microscope field.
- The server does not keep uploaded images. Your result history stays in your browser tab and holds labels only.
- The hosted API can be switched off at any time. The web app then runs in demo mode with bundled results.
- The Android app is still in its technical phase.