Harvard, Boston, MA
Muhammad Shaban , Wiem Lassoued , Kenneth Canubas , Shania Bailey , Yanling Liu , Clint Allen , Julius Strauss , James L. Gulley , Sizun Jiang , Faisal Mahmood , George Zaki , Houssein Abdul Sater
Background: Multiplex staining and imaging, a state-of-the-art technology, has revolutionized the simultaneous visualization of multiple protein markers within a single tissue sample. Various techniques have emerged to capture multiplex images with up to one hundred markers, enabling a deeper understanding of complex biological processes. The increased marker count increased the likelihood of staining and imaging failure, leading to higher resource usage in multiplex staining and imaging. We address this challenge by proposing a deep learning method and leveraging latent biological relationships between markers to accurately impute unstained protein markers. Methods: A deep learning-based marker imputation model for multiplex images (MAXIM) was developed and trained. The model’s imputation ability is evaluated at pixel and cell levels across various cancer types. Additionally, we present a comparison between imputed and actual marker images within the context of a downstream cell classification task. The MAXIM model’s interpretability is enhanced by gaining insights into the contribution of individual markers in the imputation process. Results: MAXIM was successfully trained and evaluated on a whole slide multiplex immunofluorescence (mIF) imaging datasets (14,476 images), encompassing cases from four different cancer types: Urothelial, Anal, Cervical, Head and Neck Squamous Cell Carcinoma (HNSCC). A separate MAXIM model was trained for each marker in mIF images, using the remaining markers as input. MAXIM performance was evaluated using structural similarity index (SSIM) and mean absolute error (MAE) between the imputed marker images and corresponding real marker images. MAXIM achieved high median SSIM, and low median MAE scores as well as high precision scores (AUC 0.95-0.99). Conclusions: The MAXIM’s method provides a platform with multiple potentials. First, laboratories can seamlessly train an in-house MAXIM model using images devoid of staining issues. The trained model can then be employed to accurately impute markers in multiplexed images that are marred by staining problems. Second, MAXIM can serve as a valuable tool for quality control in newly generated multiplex images, aiding in the detection of staining failures. The strong correlation between imputed and real markers in new images will be an indicator of staining integrity. In practice, MAXIM can reduce the cost and time of multiplex staining and image acquisition by accurately imputing protein markers with less staining. Third the interpretability of MAXIM provides the opportunity to uncover previously unknown latent biological relationships between different protein markers, leading to new insights in the field. Finally, the method can be scaled up for discovery of novel and clinically relevant biomarkers beneficial for offering targeted treatments in different cancer types.
Disclaimer
This material on this page is ©2024 American Society of Clinical Oncology, all rights reserved. Licensing available upon request. For more information, please contact licensing@asco.org
Abstract Disclosures
2024 ASCO Annual Meeting
First Author: Chiharu Sako
2024 ASCO Genitourinary Cancers Symposium
First Author: Keara English
2024 ASCO Gastrointestinal Cancers Symposium
First Author: Jon Stanley Heiselman
2022 ASCO Annual Meeting
First Author: Eun Kyung Park