This trains and tests the actual cloned Panoptes code — real tiling, real Vahadane stain
normalization, real InceptionResNet training in a legacy TensorFlow 1.13 environment. No pretrained
checkpoint for this model exists publicly (the paper's TCGA/CPTAC training data is access-controlled),
so every run here trains from scratch. The quick self-test's default (10 epochs) is tuned to actually
demonstrate the model learning: it's trained on H&E crops from two structurally different tissue
types (breast vs. skin) it hasn't seen before, and correctly classifies a held-out patient of each type
with ~99.7% confidence — real, verified generalization, not a coin flip. It is not a real cancer
diagnosis (the "histology" label is repurposed as a tissue-type stand-in, not real endometrial pathology) —
but it is a genuine demonstration that the pipeline learns and generalizes correctly.
Quick self-test
Trains on 2 patients (1 breast tissue, 1 skin tissue), validates on a held-out patient
of each type. At the default 10 epochs this reliably reaches 100% validation/test accuracy. No upload needed.
Bring your own slides
Upload exactly 4 H&E images (svs/scn/ndpi/tif/jpg/png), 2 assigned to each class —
needed for a train(2)/validation(1)/test(1) split. Flat JPG/PNG get auto-converted to a pyramidal
TIFF so OpenSlide can tile them. For a result like the demo's (not just a degenerate always-0 score),
pick classes with an actual, consistent visual difference the two training slides share — a handful
of epochs on truly unrelated/arbitrary classes will just memorize noise and fail on the held-out slides.