PathAI reposted this
Great writeup about our new paper from PathAI -- if you're interested in nuclear morphology and want to see it in action (beyond the paper figures), it's available to interact with via our demos on some example TCGA slides https://lnkd.in/ecetsUqr
Freelance Medical & Scientific Writer | Oncology, Immunotherapy, mRNA Therapeutics, Targeted Therapies | Genomics, Molecular Diagnostics, Biomedical Data Science | Biotech Product Marketing, Publications, Medical Affairs
Alterations in nucleus size, shape, and color are ubiquitous in cancer, but comprehensive quantification of nuclear morphology across a whole-slide histologic image remains challenging. A new open-access publication by John Abel and larger team at PathAI details the development of a pan-tissue, deep learning-based digital pathology pipeline for exhaustive nucleus detection, segmentation, and classification. I included the link to the full paper and a summary below for anyone interested. AI powered quantification of nuclear morphology in cancers enables prediction of genome instability and prognosis. https://lnkd.in/esNTTDjs Methods overview: The authors collected over 29,000 manual annotations of cell nuclei from H&E images of 21 tumor types and other diseases to train and validate an object detection and segmentation model. Their model for nucleus analysis is a Mask-RCNN-style architecture with a ResNet50 backbone pretrained on ImageNet. It was deployed on whole-slide H&E images from breast cancer (TCGA BRCA), lung adenocarcinoma (TCGA LUAD), and prostate adenocarcinoma (TCGA PRAD) cohorts. Following nuclear segmentation, the cell class of each nucleus was assigned using PathExplore models specific to each cancer type. Interpretable features describing the size, shape, stain intensity, and texture were computed for each individual nucleus. The mean and standard deviation of each feature for each cell class on each slide were used to summarize the nuclear morphology. Statistical analyses were performed to assess relationships between nuclear morphology and cancer type, genomic instability, breast cancer molecular subtype, survival, and gene expression. Results overview: The nuclear segmentation and classification model performed comparably to previously reported models. It revealed differences in nuclear morphology sufficient to distinguish between BRCA, LUAD, and PRAD. Cancer cell nuclear area was associated with increased aneuploidy score and homologous recombination deficiency across cancer types. It was also predictive of whole genome doubling. In the breast BRCA cohort, cell-type-specific nuclear morphology enabled classification of luminal A, basal-like, and HER2 molecular subtypes. Increased fibroblast nuclear area in BRCA was indicative of poor progression-free and overall survival and was associated with gene expression signatures related to extracellular matrix remodeling and anti-tumor immunity. In summary, this work highlights the power of machine learning-driven quantitative nuclear morphometry in multiple cancer types at scale. The models and resulting features have the potential to aid pathologists in discerning novel biomarkers and provide meaningful prognostic information for cancer patients.
