AI Virchow model outperforms clinical methods in cancer detection

In a recent study published in the journal Nature Medicinea large group of researchers in the United States discussed the use of the fundamental Virchow model for computational analysis of pathological reports and demonstrated its use in histopathological analysis for biomarker prediction and cell identification in seven rare and nine common cancers.

The training dataset, training algorithm, and implementation of Virchow, a fundamental model for computational pathology. oneTraining data can be described in terms of patients, cases, samples, blocks, or slides, as shown. si–HeyThe distribution of slides as a function of cancer status (si), surgery (do) and tissue type (Hey). mThe data flow during training requires the slide to be processed into tiles, which are then cut into global and local views. eat, Schematic illustration of applications of the foundation model using an accumulator model to predict transparency-level properties. GI, gastrointestinal. Study: A fundamental model for clinical-grade computational pathology and rare cancer detection

Record

Diagnosis of cancers has traditionally depended on examination of histopathological preparations of hematoxylin and eosin slides using light microscopes. Advances in digital technology and computational pathology have replaced this with computerized whole-slide images, making this form of diagnosis part of routine clinical practice.

The use of artificial intelligence (AI) in cancer diagnosis and characterization using digitized whole-slide images has also grown significantly, with initial efforts focused on improving workflows. However, recent studies have explored a subfield where artificial intelligence is used extensively to analyze whole-slide images to reveal more than just diagnostic information, including therapeutic responses and prognosis. This also reduces reliance on genomic testing and immunohistochemistry-based methods for cancer diagnosis.

About the study

In the present study, the researchers discussed the largest fundamental model developed to date, the Virchow. They demonstrated its use in predicting cancer biomarkers in a wide range of common and rare cancers.

Fundamental models are large-scale neural networks that are trained on very large datasets using self-supervised learning. These models create data representations known as embeddings, which can gather generalized data from large data sets to be applied in data-poor situations and for predictive tasks such as determining clinical outcomes, genomic changes, and therapeutic responses.

An effective fundamental model can capture broad-spectrum patterns such as tissue architecture, nuclear morphology, cellular morphology, necrosis, staining patterns, neovascularization, inflammation, and biomarker expression that can be used to predict various whole-slide image features.

Here, the researchers discussed Virchow, the most fundamental model developed to date, named after the pioneering modern pathologist Rudolf Virchow. The model has been trained on a large dataset of nearly 1.5 million whole-slide hematoxylin and eosin images obtained from one hundred thousand patients enrolled at Memorial Sloan Kettering Cancer Center (MSKCC). The dataset consists of benign and malignant tissue samples obtained from resections and biopsies of 17 tissue types.

Virchow is a vision transformer model that includes 632 million parameters. It is trained using a self-supervised algorithm that uses local and global regions of tissue tiles to generate whole-slide image embeddings that can be used for predictive tasks.

To highlight the clinical applications and utility of such a large fundamental model, the researchers used Virchow integrations generated from the large whole-slide image dataset to train a pan-cancer model and evaluate its performance in predicting common and rare cancers at the sample level in various tissues.

The study compared the performance of Virchow integrations against Phikon, UNI and CTransPath integrations and evaluated the utility of Virchow integrations in two categories. The first was the performance of the pan-cancer detection model trained using Virchow embeddings on a test dataset consisting of a mixture of datasets from MSKCC and external sources spanning seven rare and nine common cancer types. The effectiveness of Virchow integrations in predicting biomarkers was also evaluated using data from cancers such as lung, bladder, breast, and colon cancer.

Results

The study showed that Virchow integrations proved the dual value of a fundamental pathology model, being generalizable and providing training data efficiency. The pan-cancer model trained on Virchow embeddings was able to detect not only the common cancers but also the rarer histological subtypes in the test data set.

In addition, the performance of the pan-cancer model was comparable to that of clinical-grade cancer detection models and, in some rare cancers, even exceeded that of clinical models, even though it was trained on datasets with fewer tissue-specific labels.

The researchers said the model’s level of performance was particularly remarkable considering that the dataset used to train the pan-cancer model was not subjected to the subpopulation enrichment and quality control performed to train artificial intelligence models used commercially and clinically.

one,siPerformance as measured by the AUC of three clinical products compared to the pan-carcinoma model trained on Virchow embeddings, in the rare variant (one) and product test datasets (si). The pan-cancer detector, trained on embeddings of Virchow foundation models, achieves similar performance to clinical-grade products in general and outperforms them in rare cancer variants. doThe pan-cancer detector was trained on fewer labeled samples than the prostate, breast, and BLN clinical models, including a small fraction of the prostate (cyan), breast (blue), and BLN (yellow) tissue samples that these clinical models were trained on. Hey, A categorization of pan-cancer model failure patterns and four canonical examples of primary failure types. In all tables, * is used to indicate pairwise statistical significance (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, test DeLong pairwise). Error bars indicate the two-sided 95% confidence interval, calculated by DeLong's method. C., carcinoma. Inv., invasive.

conclusions

In conclusion, the study showed that a porcine model trained using Virchow integrations was able to perform comparably and often better than clinical-grade models in detecting common and rare cancers, despite being trained on datasets with fewer tissue labels.

Overall, the findings highlighted the importance and utility of fundamental models such as Virchow in applications involving a limited amount of training data, providing the basis for various clinical models in cancer pathology.