Fig. 6. Strengths and pitfalls of current methods.

Each H&E example is shown as a raw image with expert pathology annotations on some cells (Ground truth, left) and the AI cell identification (Predicted, right). Cell colours are denoted as four cell classes, green: cancer (malignant epithelial) cells; blue: lymphocytes (including plasma cells); yellow: noninflammatory stromal cells (fibroblasts and endothelial cells); white: ‘other’ cell class that included nonidentifiable cells, less abundant cells such as macrophages and chondrocytes and ‘normal’ pneumocytes. Scale bar = 50 µm. a Correct identification of cancer cells from a mesenchymal tumour (metastatic anaplastic sarcoma) in a snake (GONOXY). b A malignant spindle cell tumour from a ring-tailed lemur (LEMCAT) with a haemangiosarcoma disease, as shown, the neoplastic endothelial cells have large and rounded nucleus, which may appear morphologically similar to epithelial cancer cells, as opposed to the AI model’s own normal -stromal- endothelial cells. However, the model successfully distinguished the majority of neoplastic from stromal cells. Further complexity is in the occurrence of epithelioid haemangiosarcoma, where the cells of origin are endothelial cells but they became epithelial-like. c A challenging South American coati (NASNAS) case was diagnosed with a round-cell tumour (lymphosarcoma) where the cancer cell morphology is difficult to be recognised by an algorithm trained with epithelial cells from human lung cancer. d In the case of a chimpanzee (PANTRO) with a spindle cell sarcoma, the neoplastic fibroblasts are harder to differentiate from reactive fibroblast.