Clinical-grade computational pathology using weakly supervised deep learning on whole slide images

Campanella, Gabriele; Hanna, Matthew G.; Geneslaw, Luke; Miraflor, Allen P.; Silva, Vitor Werneck Krauss; Busam, Klaus J.; Brogi, Edi; Reuter, Victor E.; Klimstra, David S.; Fuchs, Thomas J.

doi:10.1038/s41591-019-0508-1

Cited by 1,530 publications

(1,330 citation statements)

References 32 publications

Supporting

Mentioning

1,308

Contrasting

Unclassified

Order By: Relevance

“…Whole genome duplications (WGD) occur in about 30% of solid tumors, leading to cells with a nearly tetraploid genome, likely as a result of a single failed mitosis, often preceded or succeeded by further chromosomal gains and losses 11 . WGD status could be predicted for 25 out of 27 cancer types with an average area under the receiver operating characteristic curve AUC of 0.79 (held back AUC > 0.5, false discovery rate FDR < 0.1, Figure 2a, Supplementary Table 2 ).…”

Section: Accurate Predictions Of Whole Genome Duplicationsmentioning

confidence: 99%

“…As outlined above, deep learning algorithms are capable of extracting a rich feature representation of images, which is of great value beyond the primary learning task of tissue classification. However, doing so with millions of parameters may also come at the price of overfitting to the training data 11,41 . In this context, overfitting may unearth undesirable features introduced by specimen acquisition, sample preparation, sectioning and staining, but also microscope settings as well as digital differences resulting from lossy file formats and compression parameters, the entirety of which can be hard to control for.…”

Section: Generalisation Requires Tailored Architectures and Data Augmmentioning

confidence: 99%

“…While such molecular tests are now reaching maturity in terms of automation, accuracy and reproducibility, morphological tissue assessment remains a laborious task carried out by trained histopathologists 7,8 . Computer vision and, in particular, deep convolutional neural networks (CNNs) have been shown to closely match the diagnostic accuracy of specialists and thus hold great promise to augment histopathology workflows [9][10][11] . Computational histopathology algorithms can process and cross-reference very large volumes of data, which may help pathologists to navigate and assess slides more quickly and help quantify aberrant cells and tissues.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis

Jung

Torné

et al. 2019

Preprint

105

164

View full text Add to dashboard Cite

Key points• Pan-cancer computational histopathology analysis with deep learning extracts histopathological patterns and accurately discriminates 28 cancer and 14 normal tissue types • Computational histopathology predicts whole genome duplications, focal amplifications and deletions, as well as driver gene mutations • Wide-spread correlations with gene expression indicative of immune infiltration and proliferation • Prognostic information augments conventional grading and histopathology subtyping in the majority of cancers AbstractHere we use deep transfer learning to quantify histopathological patterns across 17,396 H&E stained histopathology image slides from 28 cancer types and correlate these with underlying genomic and transcriptomic data. Pan-cancer computational histopathology (PC-CHiP) classifies the tissue origin across organ sites and provides highly accurate, spatially resolved tumor and normal distinction within a given slide. The learned computational histopathological features correlate with a large range of recurrent genetic aberrations, including whole genome duplications (WGDs), arm-level copy number gains and losses, focal amplifications and deletions as well as driver gene mutations within a range of cancer types. WGDs can be predicted in 25/27 cancer types (mean AUC=0.79) including those that were not part of model training. Similarly, we observe associations with 25% of mRNA transcript levels, which enables to learn and localise histopathological patterns of molecularly defined cell types on each slide. Lastly, we find that computational histopathology provides prognostic information augmenting histopathological subtyping and grading in the majority of cancers assessed, which pinpoints prognostically relevant areas such as necrosis or infiltrating lymphocytes on each tumour section. Taken together, these findings highlight the large potential of PC-CHiP to discover new molecular and prognostic associations, which can augment diagnostic workflows and lay out a rationale for integrating molecular and histopathological data.

show abstract

Section: Accurate Predictions Of Whole Genome Duplicationsmentioning

confidence: 99%

Section: Generalisation Requires Tailored Architectures and Data Augmmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis

Jung

Torné

et al. 2019

Preprint

105

164

View full text Add to dashboard Cite

show abstract

“…prostate biopsies, up to 75% of routine H&E slides could be screened out as being normal by a machine learning algorithm, and excluded from the pathologist's daily workload. The remaining 25% of slides could then be triaged for careful pathologist assessment, with no decrease in diagnostic sensitivity . The concept is not that far‐fetched; in fact, we have already been doing this for nearly two decades in liquid‐based Pap smear assessment .…”

Section: Digital Pathology and The Modern Pathology Laboratorymentioning

confidence: 99%

“…The remaining 25% of slides could then be triaged for careful pathologist assessment, with no decrease in diagnostic sensitivity. 9 The concept is not that far-fetched; in fact, we have already been doing this for nearly two decades in liquid-based Pap smear assessment. 10 Another similar example could be frozen section interpretation.…”

Section: Digital Pathology and The Modern Pathology Laboratorymentioning

confidence: 99%

The coming 15 years in gynaecological pathology: digitisation, artificial intelligence, and new technologies

2019

View full text Add to dashboard Cite

Surgical pathology forms the cornerstone of modern oncological medicine, owing to the wealth of clinically relevant information that can be obtained from tissue morphology. Although several ancillary testing modalities have been added to surgical pathology, the way in which we view and interpret tissue morphology has remained largely unchanged since the inception of our profession. In this review, we discuss new technological advances that promise to transform the way in which we access tissue morphology and how we use it to guide patient care.

show abstract

The Promise and Drawbacks of Federated Learning for Dermatology AI

Kose,

Rotemberg

2024

JAMA Dermatol

View full text Add to dashboard Cite

Artificial intelligence (AI) classification of images has achieved considerable promise for dermatology applications for photography, dermoscopy, and pathology for a number of years. However, a major challenge in these arenas is the curation of well-labeled data sets that can be shared publicly due to regulatory concerns as well as the common practice of using retrospective data sets that have not been prospectively consented for research use. AI algorithm performance is directly related to the size and breadth of the data used for training, so methodology that expands the potential size and diversity of inputs and evaluates the generalizability of resultant methods should be enthusiastically explored.The study in this issue of JAMA Dermatology by Haggenmüller et al 1 is an example of a proof-of-concept study that advances the field of dermatology AI by taking the challenge of barriers to public sharing of data and proposing a methodology for overcoming it by using federated learning (FL). In the study, they have curated prospectively acquired histopathologic whole slide image data from 6 hospitals for a binary classification task around invasive melanoma vs nevi. They compared the FL setting with the 2 most widely used training settings: centralized and ensemble training. In the centralized setting, the data from multiple sources are pooled together under a single data set to train a classification model. The authors rightly argue that bringing such data sets together is challenging due to data sharing and use permission issues. A potential solution to this problem is using an ensemble training approach, where models are trained separately at different centers using their own data and then combined through model ensembling. However, an important limitation of this approach is that each model is trained on a relatively small subset of the overall data set, making it susceptible to overfitting to the characteristics of the specific data site it was trained on. Using an ensemble model increases the computational complexity during inference because it consists of multiple models. In addition, ensemble models also present challenges regarding their interpretability, given that their outcomes result from a blend of predictions made by multiple individual models. The authors examine whether using an FL model may address challenges in both approaches. In FL, each contributing site independently trains a common model using only their respective data, while also communicating with a centralized server or directly with the other sites, to send and get periodic updates on the model parameters. In the study by Haggenmüller et al, 1 the authors used a FL approach, where models from each site are collected and merged at a centralized server. The merged model is obtained by

show abstract

Clinical-grade computational pathology using weakly supervised deep learning on whole slide images

Cited by 1,530 publications

References 32 publications

Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis

Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis

The coming 15 years in gynaecological pathology: digitisation, artificial intelligence, and new technologies

The Promise and Drawbacks of Federated Learning for Dermatology AI

Contact Info

Product

Resources

About