To better identify patients with early breast cancer who are at risk of relapse, the researchers from the startup Owkin and Gustave Roussy conceived a diagnostic tool that applies deep learning to digital pathology slides and clinical data. They showed that their model could predict the risk of relapse in patients with oestrogen receptor (ER)-positive, HER2-negative early breast cancer. The findings are presented by Dr. Ingrid J. Garberis of the Medical Oncology Institut Gustave Roussy in Villejuif, France during the proffered papers session on translational research at ESMO Congress 2021 (16-21 September).
Dr. Garberis told the audience that breast cancer has a favourable long-term prognosis, with an estimated average 5-year survival rate of 87%. However, 10% of patients relapse after initial treatment each year and better identification of these patients is needed. Together with Owkin, the Gustave Roussy researchers conceived a diagnostic tool that applies deep learning to whole slide images and clinical data.
In total, 1437 patients with ER-positive, HER2-negative breast cancer who were diagnosed between 2005 and 2013 were included. All patients underwent surgical resection with full follow-up and an available hematoxylin eosin-stained glass slide, which was digitised, pre-processed and cut into small patches, called tiles. These tiles were fed into the deep learning network along with survival information.
A weighted average of tile features was computed to predict a risk of relapse. Cox models based on baseline clinical variables such as age at surgery, tumour stage and size, number of positive nodes, number of nodules, surgery type, as well as on extended clinical variables that combined baseline variables with status of hormonal receptors and HER2, grade, Ki67, histological types were also considered. Performance was evaluated by using cross-validation.
Metastasis-free interval was chosen as primary endpoint. Uno’s time dependent area under the curve (AUC) was used as a metric to quantify the discrimination capability of the models.
The prediction of 5-year survival based on baseline variables yielded an AUC of 0.77. Deep learning algorithm based on whole slide image yielded an AUC of 0.77. A model based on extended variables yielded an AUC of 0.80. Combining baseline variables with deep learning algorithm resulted in an improved AUC of 0.81.
BoxPlot representing model’s performance. By integrating basic clinical variables in the deep learning (DL) model alongside to the H&E slide, prediction of 5-year distant relapse substantially improves to 0.81 AUC. Baseline variables (BV) include age, pT, number of positive lymph nodes, tumour size, number of masses and type of surgery.
© Ingrid J. Garberis
This model also predicted relapse in patients with ER-positive, HER2-negative, node negative subgroup where AUC was 0.77 and ER-positive, HER2-negative, node positive subgroup where AUC was 0.80.
The study team plans to perform validation on large independent cohorts from external healthcare centres.
Dr. Garberis concluded that deep learning applied to whole slide image could predict the risk of relapse in patients with early ER-positive, HER2-negative breast cancer. When coupled with baseline variables, their model could be a promising tool for treatment decision-making at low cost. As the next steps, the researchers plan to unravel the most predictive tiles to discover new biomarkers, and to develop novel models for prediction of relapse, as an alternative to immunohistochemistry or molecular tests.
This study was funded by a grant from the Région Ile-de-France.
Reference
1124O – Garberis IJ, Saillard C, Drubay D, et al. Prediction of distant relapse in patients with invasive breast cancer from deep learning models applied to digital pathology slides. ESMO Congress 2021 (16-21 September).