Clinical Validation of Artificial Intelligence-Powered PD-L1 Tumor Proportion Score Interpretation for Immune Checkpoint Inhibitor Response Prediction in Non-Small Cell Lung Cancer.

PURPOSE: Evaluation of PD-L1 tumor proportion score (TPS) by pathologists has been very impactful but is limited by factors such as intraobserver/interobserver bias and intratumor heterogeneity. We developed an artificial intelligence (AI)-powered analyzer to assess TPS for the prediction of immune checkpoint inhibitor (ICI) response in advanced non-small cell lung cancer (NSCLC). MATERIALS AND METHODS: The AI analyzer was trained with 393,565 tumor cells annotated by board-certified pathologists for PD-L1 expression in 802 whole-slide images (WSIs) stained by 22C3 pharmDx immunohistochemistry. The clinical performance of the analyzer was validated in an external cohort of 430 WSIs from patients with NSCLC. Three pathologists performed annotations of this external cohort, and their consensus TPS was compared with AI-based TPS. RESULTS: In comparing PD-L1 TPS assessed by AI analyzer and by pathologists, a significant positive correlation was observed (Spearman coefficient = 0.925; P < .001). The concordance of TPS between AI analyzer and pathologists according to TPS ≥50%, 1%-49%, and <1% was 85.7%, 89.3%, and 52.4%, respectively. In median progression-free survival (PFS), AI-based TPS predicted prognosis in the TPS 1%-49% or TPS <1% group better than the pathologist's reading, with the TPS ≥50% group as a reference (hazard ratio [HR], 1.49 [95% CI, 1.19 to 1.86] v HR, 1.36 [95% CI, 1.08 to 1.71] for TPS 1%-49% group, and HR, 2.38 [95% CI, 1.69 to 3.35] v HR, 1.62 [95% CI, 1.23 to 2.13] for TPS <1% group). CONCLUSION: PD-L1 TPS assessed by AI analyzer correlates with that of pathologists, with clinical performance also being comparable when referenced to PFS. The AI model can accurately predict tumor response and PFS of ICI in advanced NSCLC via assessment of PD-L1 TPS.

Artificial Intelligence-Powered Whole-Slide Image Analyzer Reveals a Distinctive Distribution of Tumor-Infiltrating Lymphocytes in Neuroendocrine Neoplasms.

Despite the importance of tumor-infiltrating lymphocytes (TIL) and PD-L1 expression to the immune checkpoint inhibitor (ICI) response, a comprehensive assessment of these biomarkers has not yet been conducted in neuroendocrine neoplasm (NEN). We collected 218 NENs from multiple organs, including 190 low/intermediate-grade NENs and 28 high-grade NENs. TIL distribution was derived from Lunit SCOPE IO, an artificial intelligence (AI)-powered hematoxylin and eosin (H&E) analyzer, as developed from 17,849 whole slide images. The proportion of intra-tumoral TIL-high cases was significantly higher in high-grade NEN (75.0% vs. 46.3%, p = 0.008). The proportion of PD-L1 combined positive score (CPS) ≥ 1 case was higher in high-grade NEN (85.7% vs. 33.2%, p < 0.001). The PD-L1 CPS ≥ 1 group showed higher intra-tumoral, stromal, and combined TIL densities, compared to the CPS < 1 group (7.13 vs. 2.95, p < 0.001; 200.9 vs. 120.5, p < 0.001; 86.7 vs. 56.1, p = 0.004). A significant correlation was observed between TIL density and PD-L1 CPS (r = 0.37, p < 0.001 for intra-tumoral TIL; r = 0.24, p = 0.002 for stromal TIL and combined TIL). AI-powered TIL analysis reveals that intra-tumoral TIL density is significantly higher in high-grade NEN, and PD-L1 CPS has a positive correlation with TIL densities, thus showing its value as predictive biomarkers for ICI response in NEN.

Artificial intelligence-powered programmed death ligand 1 analyser reduces interobserver variation in tumour proportion score for non-small cell lung cancer with better prediction of immunotherapy response.

BACKGROUND: Manual evaluation of programmed death ligand 1 (PD-L1) tumour proportion score (TPS) by pathologists is associated with interobserver bias. OBJECTIVE: This study explored the role of artificial intelligence (AI)-powered TPS analyser in minimisation of interobserver variation and enhancement of therapeutic response prediction. METHODS: A prototype model of an AI-powered TPS analyser was developed with a total of 802 non-small cell lung cancer (NSCLC) whole-slide images. Three independent board-certified pathologists labelled PD-L1 TPS in an external cohort of 479 NSCLC slides. For cases of disagreement between each pathologist and the AI model, the pathologists were asked to revise the TPS grade (<1%, 1%-49% and ≥50%) with AI assistance. The concordance rates among the pathologists with or without AI assistance and the effect of the AI-assisted revision on clinical outcome upon immune checkpoint inhibitor (ICI) treatment were evaluated. RESULTS: Without AI assistance, pathologists concordantly classified TPS in 81.4% of the cases. They revised their initial interpretation by using the AI model for the disagreement cases between the pathologist and the AI model (N = 91, 93 and 107 for each pathologist). The overall concordance rate among the pathologists was increased to 90.2% after the AI assistance (P < 0.001). A reduction in hazard ratio for overall survival and progression-free survival upon ICI treatment was identified in the TPS subgroups after the AI-assisted TPS revision. CONCLUSION: The AI-powered TPS analyser assistance improves the pathologists' consensus of reading and prediction of the therapeutic response, raising a possibility of standardised approach for the accurate interpretation.

Artificial Intelligence-Powered Hematoxylin and Eosin Analyzer Reveals Distinct Immunologic and Mutational Profiles among Immune Phenotypes in Non-Small-Cell Lung Cancer.

The tumor microenvironment can be classified into three immune phenotypes: inflamed, immune excluded, and immune-desert. Immunotherapy efficacy has been shown to vary by phenotype; yet, the mechanisms are poorly understood and demand further investigation. This study unveils the mechanisms using an artificial intelligence-powered software called Lunit SCOPE. Artificial intelligence was used to classify 965 samples of non-small-cell lung carcinoma from The Cancer Genome Atlas into the three immune phenotypes. The immune and mutational profiles that shape each phenotype using xCell, gene set enrichment analysis with RNA-sequencing data, and cBioportal were described. In the inflamed subtype, which showed higher cytolytic score, the enriched pathways were generally associated with immune response and immune-related cell types were highly expressed. In the immune excluded subtype, enriched glycolysis, fatty acid, and cholesterol metabolism pathways were observed. The KRAS mutation, BRAF mutation, and MET splicing variant were mostly observed in the inflamed subtype. The two prominent mutations found in the immune excluded subtype were EGFR and PIK3CA mutations. This study is the first to report the distinct immunologic and mutational landscapes of immune phenotypes, and demonstrates the biological relevance of the classification. In light of these findings, the study offers insights into potential treatment options tailored to each immune phenotype.

Artificial Intelligence-Powered Spatial Analysis of Tumor-Infiltrating Lymphocytes as Complementary Biomarker for Immune Checkpoint Inhibition in Non-Small-Cell Lung Cancer.

PURPOSE: Biomarkers on the basis of tumor-infiltrating lymphocytes (TIL) are potentially valuable in predicting the effectiveness of immune checkpoint inhibitors (ICI). However, clinical application remains challenging because of methodologic limitations and laborious process involved in spatial analysis of TIL distribution in whole-slide images (WSI). METHODS: We have developed an artificial intelligence (AI)-powered WSI analyzer of TIL in the tumor microenvironment that can define three immune phenotypes (IPs): inflamed, immune-excluded, and immune-desert. These IPs were correlated with tumor response to ICI and survival in two independent cohorts of patients with advanced non-small-cell lung cancer (NSCLC). RESULTS: Inflamed IP correlated with enrichment in local immune cytolytic activity, higher response rate, and prolonged progression-free survival compared with patients with immune-excluded or immune-desert phenotypes. At the WSI level, there was significant positive correlation between tumor proportion score (TPS) as determined by the AI model and control TPS analyzed by pathologists (P < .001). Overall, 44.0% of tumors were inflamed, 37.1% were immune-excluded, and 18.9% were immune-desert. Incidence of inflamed IP in patients with programmed death ligand-1 TPS at < 1%, 1%-49%, and ≥ 50% was 31.7%, 42.5%, and 56.8%, respectively. Median progression-free survival and overall survival were, respectively, 4.1 months and 24.8 months with inflamed IP, 2.2 months and 14.0 months with immune-excluded IP, and 2.4 months and 10.6 months with immune-desert IP. CONCLUSION: The AI-powered spatial analysis of TIL correlated with tumor response and progression-free survival of ICI in advanced NSCLC. This is potentially a supplementary biomarker to TPS as determined by a pathologist.

Tumor-infiltrating lymphocyte enrichment predicted by CT radiomics analysis is associated with clinical outcomes of non-small cell lung cancer patients receiving immune checkpoint inhibitors.

Background: Enrichment of tumor-infiltrating lymphocytes (TIL) in the tumor microenvironment (TME) is a reliable biomarker of immune checkpoint inhibitors (ICI) in non-small cell lung cancer (NSCLC). Phenotyping through computed tomography (CT) radiomics has the overcome the limitations of tissue-based assessment, including for TIL analysis. Here, we assess TIL enrichment objectively using an artificial intelligence-powered TIL analysis in hematoxylin and eosin (H&E) image and analyze its association with quantitative radiomic features (RFs). Clinical significance of the selected RFs is then validated in the independent NSCLC patients who received ICI. Methods: In the training cohort containing both tumor tissue samples and corresponding CT images obtained within 1 month, we extracted 86 RFs from the CT images. The TIL enrichment score (TILes) was defined as the fraction of tissue area with high intra-tumoral or stromal TIL density divided by the whole TME area, as measured on an H&E slide. From the corresponding CT images, the least absolute shrinkage and selection operator model was then developed using features that were significantly associated with TIL enrichment. The CT model was applied to CT images from the validation cohort, which included NSCLC patients who received ICI monotherapy. Results: A total of 220 NSCLC samples were included in the training cohort. After filtering the RFs, two features, gray level variance (coefficient 1.71 x 10-3) and large area low gray level emphasis (coefficient -2.48 x 10-5), were included in the model. The two features were both computed from the size-zone matrix, which has strength in reflecting intralesional texture heterogeneity. In the validation cohort, the patients with high predicted TILes (≥ median) had significantly prolonged progression-free survival compared to those with low predicted TILes (median 4.0 months [95% CI 2.2-5.7] versus 2.1 months [95% CI 1.6-3.1], p = 0.002). Patients who experienced a response to ICI or stable disease with ICI had higher predicted TILes compared with the patients who experienced progressive disease as the best response (p = 0.001, p = 0.036, respectively). Predicted TILes was significantly associated with progression-free survival independent of PD-L1 status. Conclusions: In this CT radiomics model, predicted TILes was significantly associated with ICI outcomes in NSCLC patients. Analyzing TME through radiomics may overcome the limitations of tissue-based analysis and assist clinical decisions regarding ICI.

Deep learning from HE slides predicts the clinical benefit from adjuvant chemotherapy in hormone receptor-positive breast cancer patients.

We hypothesized that a deep-learning algorithm using HE images might be capable of predicting the benefits of adjuvant chemotherapy in cancer patients. HE slides were retrospectively collected from 1343 de-identified breast cancer patients at the Samsung Medical Center and used to develop the Lunit SCOPE algorithm. Lunit SCOPE was trained to predict the recurrence using the 21-gene assay (Oncotype DX) and histological parameters. The risk prediction model predicted the Oncotype DX score > 25 and the recurrence survival of the prognosis validation cohort and TCGA cohorts. The most important predictive variable was the mitotic cells in the cancer epithelium. Of the 363 patients who did not receive adjuvant therapy, 104 predicted high risk had a significantly lower survival rate. The top-300 genes highly correlated with the predicted risk were enriched for cell cycle, nuclear division, and cell division. From the Oncotype DX genes, the predicted risk was positively correlated with proliferation-associated genes and negatively correlated with prognostic genes from the estrogen category. An integrative analysis using Lunit SCOPE predicted the risk of cancer recurrence and the early-stage hormone receptor-positive breast cancer patients who would benefit from adjuvant chemotherapy.

Accuracy and efficiency of an artificial intelligence tool when counting breast mitoses.

BACKGROUND: The mitotic count in breast carcinoma is an important prognostic marker. Unfortunately substantial inter- and intra-laboratory variation exists when pathologists manually count mitotic figures. Artificial intelligence (AI) coupled with whole slide imaging offers a potential solution to this problem. The aim of this study was to accordingly critique an AI tool developed to quantify mitotic figures in whole slide images of invasive breast ductal carcinoma. METHODS: A representative H&E slide from 320 breast invasive ductal carcinoma cases was scanned at 40x magnification. Ten expert pathologists from two academic medical centers labeled mitotic figures in whole slide images to train and validate an AI algorithm to detect and count mitoses. Thereafter, 24 readers of varying expertise were asked to count mitotic figures with and without AI support in 140 high-power fields derived from a separate dataset. Their accuracy and efficiency of performing these tasks were calculated and statistical comparisons performed. RESULTS: For each experience level the accuracy, precision and sensitivity of counting mitoses by users improved with AI support. There were 21 readers (87.5%) that identified more mitoses using AI support and 13 reviewers (54.2%) that decreased the quantity of falsely flagged mitoses with AI. More time was spent on this task for most participants when not provided with AI support. AI assistance resulted in an overall time savings of 27.8%. CONCLUSIONS: This study demonstrates that pathology end-users were more accurate and efficient at quantifying mitotic figures in digital images of invasive breast carcinoma with the aid of AI. Higher inter-pathologist agreement with AI assistance suggests that such algorithms can also help standardize practice. Not surprisingly, there is much enthusiasm in pathology regarding the prospect of using AI in routine practice to perform mundane tasks such as counting mitoses.

Predicting breast tumor proliferation from whole-slide images: The TUPAC16 challenge.

Tumor proliferation is an important biomarker indicative of the prognosis of breast cancer patients. Assessment of tumor proliferation in a clinical setting is a highly subjective and labor-intensive task. Previous efforts to automate tumor proliferation assessment by image analysis only focused on mitosis detection in predefined tumor regions. However, in a real-world scenario, automatic mitosis detection should be performed in whole-slide images (WSIs) and an automatic method should be able to produce a tumor proliferation score given a WSI as input. To address this, we organized the TUmor Proliferation Assessment Challenge 2016 (TUPAC16) on prediction of tumor proliferation scores from WSIs. The challenge dataset consisted of 500 training and 321 testing breast cancer histopathology WSIs. In order to ensure fair and independent evaluation, only the ground truth for the training dataset was provided to the challenge participants. The first task of the challenge was to predict mitotic scores, i.e., to reproduce the manual method of assessing tumor proliferation by a pathologist. The second task was to predict the gene expression based PAM50 proliferation scores from the WSI. The best performing automatic method for the first task achieved a quadratic-weighted Cohen's kappa score of κ = 0.567, 95% CI [0.464, 0.671] between the predicted scores and the ground truth. For the second task, the predictions of the top method had a Spearman's correlation coefficient of r = 0.617, 95% CI [0.581 0.651] with the ground truth. This was the first comparison study that investigated tumor proliferation assessment from WSIs. The achieved results are promising given the difficulty of the tasks and weakly-labeled nature of the ground truth. However, further research is needed to improve the practical utility of image analysis methods for this task.

From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The CAMELYON17 Challenge.

Automated detection of cancer metastases in lymph nodes has the potential to improve the assessment of prognosis for patients. To enable fair comparison between the algorithms for this purpose, we set up the CAMELYON17 challenge in conjunction with the IEEE International Symposium on Biomedical Imaging 2017 Conference in Melbourne. Over 300 participants registered on the challenge website, of which 23 teams submitted a total of 37 algorithms before the initial deadline. Participants were provided with 899 whole-slide images (WSIs) for developing their algorithms. The developed algorithms were evaluated based on the test set encompassing 100 patients and 500 WSIs. The evaluation metric used was a quadratic weighted Cohen's kappa. We discuss the algorithmic details of the 10 best pre-conference and two post-conference submissions. All these participants used convolutional neural networks in combination with pre- and postprocessing steps. Algorithms differed mostly in neural network architecture, training strategy, and pre- and postprocessing methodology. Overall, the kappa metric ranged from 0.89 to -0.13 across all submissions. The best results were obtained with pre-trained architectures such as ResNet. Confusion matrix analysis revealed that all participants struggled with reliably identifying isolated tumor cells, the smallest type of metastasis, with detection rates below 40%. Qualitative inspection of the results of the top participants showed categories of false positives, such as nerves or contamination, which could be targeted for further optimization. Last, we show that simple combinations of the top algorithms result in higher kappa metric values than any algorithm individually, with 0.93 for the best combination.