Project 1
Pan-Species Pathology Atlas
Infrastructure Building and Data Sharing
1. Digital Biobank
2. Curated annotations
The availability of digital tumour samples is limited for veterinary studies. We faced that constraint during our work.
With this inter-institutional effort, we aim to publicly release for academic use a rich cohort of H&E digital slides from 20 different species.
Check the resources here:
During the development of the foundational project of this network, we curated more than 40 thousand single-cell annotations.
We aim to publicly share this resource, hoping to contribute to the development of new tools in comparative computational pathology. <Synapse project>
3. Guidelines
We want to share our experience based on our recent work on animal tissue samples.
We detail guidance on the curation of high-quality images involving the process of digitalisation, quality control, running an available AI pipeline, comparing predictions/pathologists' annotations, and implementing a new metric to better understand the model's performance from cell morphology and staining patterns.
Guidelines for Comparative Computational Pathology
(work in progress, last update October 2022)
Through this network, we aim to systematise our efforts on comparative digital pathology, the long-term goal is to build a platform for image-based approaches certified by veterinary pathologists taking as inspiration the International Immuno-Oncology Biomarker Working Group on Breast Cancer. We are still in the process of seeking funding in order to keep developing these guidelines and protocols together with the network and working group (contact us!).
1. Slide digitalisation and quality control
Tumour samples are obtained through tissue biopsies from surgery or routine postmortem examinations from animals that were i) examined immediately after euthanasia or ii) stored at 4 degrees Celsius and examined within two days from death. Lesions were excised, fixed in 10% neutral buffered formalin solution, paraffin-embedded, sectioned, and stained with H&E for analysis. All slides were scanned using NanoZoomer S210 digital slide scanner (C13239-01) and NanoZoomer digital pathology system v.3.1.7 (Hamamatsu) at 40X (228 nm/pixel resolution). During the scanning, it is recommended to examine each sample carefully, cleaning marks on the slide or coverslip. Also, during the scanning, set enough focus points to have cells in focus and avoid misclassification. Exclusion criteria defined for quality control include the presence of hemorrhage, ample necrotic tissue, the lack of tumour components, and the presence of high amounts of melanin/pigments in the tissue samples hindering the correct identification of individual cells.
2. Running the AI human-lung model
We applied a deep learning-based single-cell analysis pipeline designed and developed for human lung tumour specimens. Firstly, all viable H&E tumour tissue areas are segmented. Secondly, within the segmented tissue image, a spatially-constrained convolutional neural network predicts for each pixel the probability that it belongs to the centre of a nucleus; cell nuclei were then detected from the probability map obtained from the deep network. Lastly, each identified cell was classified into one out of four cell classes: cancer (malignant epithelial) cells, lymphocytes (including plasma cells), noninflammatory stromal cells (fibroblasts and endothelial cells) and an ‘other’ cell type that included non-identifiable cells, less abundant cells such as macrophages and chondrocytes and ‘normal’ pneumocytes and bronchial epithelial cells.
3. Quantitative validation with pathologists' annotations
An essential step of automated pipelines is the quantitative evaluation of their performance. HEre, we contrasted the predictions of the model at single-cell resolution with 41,567 pathological single-cell annotations specifically curated for veterinary computational pathology. The contrast at single-cell resolution enables the computation of standard statistics for reporting model performance (sensitivity, specificity, accuracy, balanced accuracy, F1-score, AUCROC)
4. Performance explainability: computation of morphospace overlap
We developed the morphospace overlap as a biologically-clearer way to understand the model's performance. To explore the morphological similarity between human and non-human cells and to explain the transferability of the human-lung model, we visualised and quantified the morphological space of ~32,000 cells annotated by expert pathologists. The feature space consisted in 27 features characterising the morphology and the staining pattern of each cell. We apply tSNE dimension reduction to 2D and ecological metric to measure overlap between, in this case, morphological spaces of animal and human cells.