Abstract 1220P
Background
Determination of programmed death-ligand (PD-L1) status in triple-negative breast cancer (BC) by immunohistochemical (IHC) assay represents has prognostic significance in association with immune checkpoint inhibitor (ICI) therapies. However, the evaluation of PD-L1 SP142 remains challenging for pathologists since the associated tumor microenvironment complicates the immune cell (IC) scoring. Moreover, PD-L1 distribution in tumor could also lead to a risk of inaccurate patient allocation for ICI. To accurately diagnose PD-L1 status as IC score, a computer-aided artificial intelligence (AI) system is required to precisely quantify PD-L1 expression in two-dimensional and three-dimensional (3D) pathological images.
Methods
In this study, we proposed a computer-aided AI system for digital images of IHC and 3D immunofluorescence (IF) assays. This system consisted of machine learning algorithms for tumor/IC recognition, infiltration identification, and PD-L1 expression detection. Tissue clearing technology with IF and confocal microscopy were utilized to detect the spatial distribution of PD-L1 in BC specimens.
Results
The designed computer-aided AI system was based on a tumor segmentation model with a >80% accuracy and an IC recognition model with a 90% accuracy. For the IHC digital pathology, the computer-aided AI system achieved a concordance rate of 83% in comparison with PD-L1 IC interpretation by experienced pathologists. For 3D digital images, the system achieved a concordance rate of 90% with traditional pathological diagnosis. This system could serve as an integrated plugin onto existing open-source platforms. Appling the system for 3D analysis, the spatial distribution of PD-L1 showed heterogeneity across different layers. Notably, 30% cases crossed the 1% cutoff along the 3D layers, which is related to admission of immunotherapy.
Conclusions
The computer-aided AI system could improve the PD-L1 diagnosis from IHC to 3D digital images, which provides accurate prognostic significance for precision medicine.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.
Resources from the same session
1203P - Role of tumor markers before or during chemotherapy for poorly differentiated neuroendocrine carcinomas of the digestive system: An exploratory analysis of JCOG1213
Presenter: Tomoyuki Satake
Session: Poster session 14
1204TiP - Iadademstat in combination with paclitaxel in relapsed/refractory small cell lung carcinoma (SCLC) and extrapulmonary high grade neuroendocrine carcinoma (NEC)
Presenter: Neel Belani
Session: Poster session 14
1212P - Predictive value of a near-term prediction model for severe irAEs in cancer treatment with ICIs
Presenter: Jun Zhao
Session: Poster session 14
1213P - HRD complete: A novel NGS assay for detecting homologous recombination repair (HRR) gene alterations in prostate cancer
Presenter: Xin Ye
Session: Poster session 14
1214P - A novel machine learning based method to detect homozygous deletion of homologous recombination repair (HRR) genes in prostate cancer
Presenter: Jianqing Wang
Session: Poster session 14
1215P - Comparative analysis of cfDNA liquid biopsy and tumor-based next-generation sequencing (NGS) approaches
Presenter: Anastasiya Yudina
Session: Poster session 14
1216P - A spectroscopic liquid biopsy for the earlier detection of multiple cancer types
Presenter: Matthew Baker
Session: Poster session 14
1217P - Clinical evaluation of a CE-IVD liquid biopsy pan cancer genomic profiling test
Presenter: Timothy Crook
Session: Poster session 14
1218P - Exploring cancer care pathways in seven European countries: Identifying obstacles and opportunities for the role of artificial intelligence
Presenter: Shereen Nabhani
Session: Poster session 14