Artificial Intelligence-Assisted Score Analysis for Predicting the Expression of the Immunotherapy Biomarker PD-L1 in Lung Cancer

Abstract

Programmed cell death ligand 1 (PD-L1) is a critical biomarker for predicting the response to immunotherapy. However, traditional quantitative evaluation of PD-L1 expression using immunohistochemistry staining remains challenging for pathologists. Here we developed a deep learning (DL)-based artificial intelligence (AI) model to automatically analyze the immunohistochemical expression of PD-L1 in lung cancer patients. A total of 1,288 patients with lung cancer were included in the study. The diagnostic ability of three different AI models (M1, M2, and M3) was assessed in both PD-L1 (22C3) and PD-L1 (SP263) assays. M2 and M3 showed improved performance in the evaluation of PD-L1 expression in the PD-L1 (22C3) assay, especially at 1% cutoff. Highly accurate performance in the PD-L1 (SP263) was also achieved, with accuracy and specificity of 96.4 and 96.8% in both M2 and M3, respectively. Moreover, the diagnostic results of these three AI-assisted models were highly consistent with those from the pathologist. Similar performances of M1, M2, and M3 in the 22C3 dataset were also obtained in lung adenocarcinoma and lung squamous cell carcinoma in both sampling methods. In conclusion, these results suggest that AI-assisted diagnostic models in PD-L1 expression are a promising tool for improving the efficiency of clinical pathologists.

Keywords: AI; NSCLC; PD-L1; automated scoring; pathological diagnosis.

Figure 1

Annotation dataset for tumor detection. Patch datasets and cell datasets annotated in the whole-slide images of PD-L1 staining. Both 256 × 256 patch size and 128 × 128 patch size were included in the patch datasets. In the cell datasets, PD-L1-positive tumor cells, PD-L1-positive immune cells, and PD-L1-negative tumor cells were labeled with different colors.

Figure 2

Flow chart of the study. Flow chart of tumor proportion score assessed with artificial intelligence-based diagnostic models in the pathological sections of lung cancer tissue samples stained with PD-L1 (22C3).

Figure 3

Whole-slide image inference workflow. A whole-slide image was analyzed with different artificial intelligence model-based workflows (M1, M2, and M3). The detailed information of these three workflows is shown here.

Figure 4

Deep learning (DL) model performance evaluation in the PD-L1 (22C3) and PD-L1 (SP263) assays. (A–C) Histograms of DL model performance with PD-L1 (22C3) assay test. (D–F) Histograms of DL model performance with PD-L1 (SP263) assay test. Tumor proportion score cutoff values of 1% (A, D) and 50% (B, E). Kappa score analysis (C, F).

Figure 5

Examples of tumor detection and PD-L1 calculation. (A) Example of whole-slide image analysis for tumor recognition with the 256 patch. (B) Examples of cells detected by the YOLO model: PD-L1-negative tumor cell (blue), PD-L1-positive immune cell (green), and PD-L1-positive tumor cell (red).

Figure 6

PD-L1-positive immune cells patch filter module. (A) Predicted tumor and immune patch annotated with blue and green squares, respectively. The predicted PD-L1-negative tumor cells and PD-L1-positive tumor cells are indicated with blue and red dots, respectively. (B) Performance of the immune filter module in M1 (left) and M2 (right). These pie charts show the percentage of false positive slides with CPS <1 and ≥1.

Figure 7

Deep learning (DL) model performance in different tumor types and surgical methods. (A–D) Histograms of DL model performance of lung adenocarcinoma (A, C) and lung squamous cell carcinoma (B, D) at the cutoff of 1% (A, B) and 50% (C, D). (D–F) Histograms of DL model performance of samples from surgery (E, G) and needle biopsy (F, H) at the cutoff 1% (E, F) and 50% (G, H).