AI pathology interpretation (computational pathology)

Computational pathology applies deep learning to gigapixel whole-slide images (WSIs) of stained tissue, up to 100,000 by 100,000 pixels, to automate the microscopic reading that pathologists perform manually. As the global pathologist workforce shrinks against rising biopsy volumes, AI decision- support systems take over routine tasks (mitosis counting, lymph-node metastasis search, Gleason grading of prostate cancer) and compress time to diagnosis from days to minutes, while standardizing reports and flagging suspicious foci as heatmaps. A second frontier predicts hidden molecular genetics (microsatellite instability MSI, BRAF, EGFR mutation status, PD-L1 expression) directly from cell morphology, sparing a sequencing round. The US leads commercialization under the FDA De Novo and Breakthrough Device pathways, the EU builds the 3-million-slide IMI BIGPICTURE training repository under CE-MDR, and China scales cloud telepathology for cervical, oesophageal and lung-cancer screening. The six organizations span AI software (Paige.AI, PathAI, Owkin), scanner and cloud workflow (Philips), and population-scale screening AI (Tencent Medical AI Lab, Infervision).

Key directions of AI pathology interpretation:

1. Cancer detection and grading: CNN and vision-transformer models localise micro-metastases and grade tumours (Gleason score) with pathologist-level concordance.
2. Predictive molecular pathology: morphology-only inference of MSI, BRAF, EGFR and PD-L1 status, avoiding a separate sequencing or IHC round.
3. Attention multiple-instance learning: whole-slide classification trained from only a slide-level diagnosis label, with no per-cell annotation.
4. Digital pathology workflow and telepathology: high-throughput scanners and cloud LIS integration stream cases to remote pathologists.

Sectoral value chain#

[glass slide] ──> [WSI scanner] ──> [tiling & feature extraction (ViT)]
                                                          │
                                          (attention-MIL aggregation)
                                                          │
                                                          ▼
[pathologist sign-out in LIS] <─── [tumour heatmap & report]

Value chain levels#

Cross-cutting technologies of the sector:

• whole-slide-imaging: gigapixel digital slides (up to 100,000x100,000 px at x40) scanned at 0.25 micron/pixel for nuclear-chromatin review.
• attention-multiple-instance-learning: weakly supervised slide classification that learns from a single slide-level label by attention-weighting thousands of patch embeddings.
• self-supervised-vit-encoders: contrastive self-supervised vision transformers that produce reusable 1024-dim patch embeddings without manual annotation.

The stack pairs high-resolution digital pathology with self-supervised deep learning and weakly supervised whole-slide classification.

1. Whole-slide imaging and preprocessing:
   • an H&E-stained biopsy slide is scanned on a Leica Aperio GT 450 or Philips IntelliSite at x40 (0.25 micron/pixel) into a 1-3 GB .svs file; OTSU thresholding segments tissue from background, and the tissue is tiled into 512x512 patches for GPU processing.
2. Self-supervised vision-transformer encoding:
   • a contrastive self-supervised vision transformer (ViT) pretrained on millions of unlabelled patches maps each 512x512 tile to a 1024-dimensional embedding, capturing nuclear-chromatin texture without manual annotation.
3. Attention-MIL aggregation and heatmap:
   • an attention pooling head weights each patch embedding to form a slide-level vector, from which classifiers output tumour probability and molecular-biomarker predictions, and a pixel-level heatmap is overlaid on the slide for the pathologist to review in QuPath or Aperio ImageScope.