Assessing Deep Learning Models in Breast Cancer Molecular Marker Prediction: A Systematic Review and Meta-Analysis

Abstract

The advancements in artificial intelligence (AI) and its incorporation into clinical care have improved the prognosis, diagnosis and treatment to an extent. Whole slide imaging and multi-omics data analysis sum up into a promising new sub-specialty of computation pathology. Pathology of cancer need quick diagnosis to initiate intervention and AI has gained much importance in this field. This paper aimed to evaluate the diagnostic accuracy of different deep learning models for predicting molecular markers of breast cancer.  This study was conducted by following the Preferred Reporting Items for Systematic Review and Meta-Analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA) guidelines. We searched the research articles according to research aims from PubMed, EMBASE and Ovid MEDLINE. For assessment of risk bias, “Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2)” was used that collect application concerns in different areas.   The RevMan version 5.4.0 was used for pooled analysis. Sensitivity and specificity were calculated with 95% confidence interval (CIs) for pooling the effect size. The included 9 studies have data of specificity and sensitivity for diagnostic method that help in the detection or interpretation of breast cancer risk or biomarkers through DP model (CNN). The deep learning model showed a generally good accuracy scores for detection of breast cancer through imaging. The area under the SROC curve was 0.940.  In contrast to the general detection, DL seems to be more sensitive and specified in diagnosing key molecular biomarkers (PR, ER, HER2 and Ki67) among BC patients. Also, the paper provides links to functional code repositories and ends with the exploration of limitations and potential of deep learning-based diagnostic systems.