Development and validation of a hierarchical approach for lymphoma classification using immunohistochemical markers.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
BACKGROUND: Accurate lymphoma classification is critical for effective treatment and immunohistochemistry is a cost-effective and time-saving approach. Although several machine learning algorithms showed effective results, they focused on a specific task of classification but not the whole classification workflow, thus impractical to be applied in clinical settings. Thus, we aim to develop an effective and economic machine learning-assisted system that can streamline the lymphoma differential diagnostic workflow using EBER in situ hybridization and immunohistochemical markers. METHODS: We included pathological reports diagnosed as lymphomas from two cancer centers (Sun Yat-sen University Cancer Center and Peking University Cancer Hospital & Institute). We proposed a hierarchical approach that mimicked the human diagnostic process and employed simplified panels of markers to perform a series of interpretable classification. The diagnostic accuracy for lymphoma pathological subtypes and the markers saving ratio were investigated in both temporal independent population and external medical center. RESULTS: A total of 14,927 patients and corresponding immunohistochemical results from two cancer centers were included. The proposed system had high discriminative ability for differentiating lymphoma pathological subtypes (measured by mean AUC in three validation cohorts, non-Hodgkin and Hodgkin lymphoma: 0.959; non-Hodgkin subtypes: 0.983; B-lymphoma subtypes: 0.868; T-lymphoma subtypes: 0.962; DLBCL subtypes: 0.957). In addition, the system's well selected characteristics can contribute to the development of agreement on panels of markers for differential diagnosis and help minimize cost of immunohistochemical marker techniques (measured by marker saving ratio compared to real clinical settings, internal primary-stage cohort: 16.45% saved, p < 0.001; internal later-stage cohort: 21.73% saved, p < 0.001; external cohort: 3.67% saved, p < 0.001). CONCLUSIONS: Machine learning-based hierarchical system using EBER in situ hybridization and IHC markers was developed, which could streamline the workflow by sequentially determining each lymphoma pathological subtype. The proposed system proved to be effective and cost-saving in independent and external validation, thus could be adopted affordably in future clinical practice.