Abstract PD6-02: Digital MammaPrint and BluePrint using machine learning and whole slide imaging

Autor: Belma Dodgas, Laura J. van't Veer, Ran Godrich, Julian Viret, Jorge S. Reis-Filho, Diederik Wehkamp, Adam Casson, Leo Grady, Donghun Lee, William Audeh, Annuska M. Glas, Christopher Kanan, Leonie J. M. J. Delahaye, Thomas J. Fuchs, Jeroen Mollink, Anke T. Witteveen, Matthew Lee, Hugo M. Horlings
Rok vydání: 2021
Předmět:
Zdroj: Cancer Research. 81:PD6-02
ISSN: 1538-7445
0008-5472
DOI: 10.1158/1538-7445.sabcs20-pd6-02
Popis: Background: The rise of whole slide imaging systems has enabled pathologists to remotely view cases in high resolution to diagnose cancer, and efficiently archive images . The advent of machine learning techniques and their application in digital pathology have facilitated the identification of histological patterns for effective diagnosis of disease. Deep learning-based solutions have been developed to detect and recognize cancer types, and automatically grade and stage tumors through evaluation of pathological features and patterns. However, no image-based solutions have been able to replicate a multivariate test, such as a risk -of recurrence assay for early stage breast cancer. Traditionally, subtle differences in gene expression have been measured by microarray or NGS, however these changes may also be recognized phenotypically from hematoxylin and eosin stained (H&E) slides by digital biomarkers developed using novel machine learning techniques. The large repository of images with MammaPrint and BluePrint results may enable us to develop digital MammaPrint and digital BluePrint biomarkers that predict the risk of distant recurrences and the molecular subtypes of a tumor sample using only H&E stained digitized tumor slides. Methods: Using over70,000 H&E images of early stage breast cancer patients in combination with machine learning techniques, digital versions of MammaPrint and BluePrint were developed. In total 20,000 images were used for feasibility and algorithm optimization, another 50,000 images were used for further finetuning. MammaPrint indices and BluePrint scores and categorical results were used to train the system. After the algorithms were optimized, they were locked and validated in an independent set of 5000 H&E stained images. The MammaPrint and BluePrint predictions were compared to the original MammaPrint and BluePrint results obtained from the microarray assay. The finalized and locked algorithms were further validated for precision and reproducibility in a large data set of xx Images and processed multiple times. Multicenter clinical validation was performed in H&E stained images of multiple series with long term follow up (tbd), totaling ##k images. In this ##K cohort of patients, #% were HR+/HER2-, #% were clinically HER2+ and x% were triple negative. Results: Using an independent dataset of 5000 samples, we compared the MammaPrint and BluePrint predictions obtained from the H&E slides to the traditional versions of MammaPrint and BluePrint based on a microarray. The binary performance of the digital MammaPrint had a concordance of xx% (with an AUC of xx%),%NPA and %PPA when compared to the traditional MammaPrint high-low classification. For digital BluePrint, the system had a concordance of xx% and a classification accuracy of xx%. The analytical performance showed a precision and reproducibility of xx% In a multicenter clinical validation the DRFI was xx% in dMP low risk and yy% in dMP high risk groups. The xxx dataset the performance was similar to the microarray and better than compared to clinical parameters. Analyses will be available by the Placeholder Abstract deadline. Conclusions: The combination of machine learning and digital pathology has enabled development of rapid and highly accurate in silico versions of MammaPrint and BluePrint. Implementation of digital H&E based risk of recurrence and molecular subtyping could enable preservation of valuable diagnostic tissue, faster turnaround time for test results, and a more cost effective approach to treatment planning tools, especially in countries that do not allow send out of human tissue and this adoption of risk scoring is low. Final conclusions tbd Citation Format: Annuska M Glas, Jorge S Reis-Filho, Diederik Wehkamp, Belma Dodgas, Leonie Delahaye, Ran Godrich, Jeroen Mollink, Adam Casson, Anke Witteveen, Julian Viret, Donghun Lee, Matthew Lee, Hugo Horlings, Leo Grady, Thomas Fuchs, William Audeh, Christopher Kanan, Laura J van't Veer. Digital MammaPrint and BluePrint using machine learning and whole slide imaging [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PD6-02.
Databáze: OpenAIRE
Externí odkaz: