Overcoming the Interobserver Variability in Lung Adenocarcinoma Subtyping: A Clustering Approach to Establish a Ground Truth for Downstream Applications

Establishment of artificial intelligence model for precise histological subtyping of lung adenocarcinoma and its application to quantitative and spatial analysis

BACKGROUND

The histological subtype of lung adenocarcinoma is a major prognostic factor. We developed a new artificial intelligence model to classify lung adenocarcinoma images into seven histological subtypes and adopted the model for whole-slide images to investigate the relationship between the distribution of histological subtypes and clinicopathological factors.

METHODS

Using histological subtype images, which are typical for pathologists, we trained and validated an artificial intelligence model. Then, the model was applied to whole-slide images of resected lung adenocarcinoma specimens from 147 cases.

RESULT

The model achieved an accuracy of 99.7% in training sets and 90.4% in validation sets consisting of typical tiles of histological subtyping for pathologists. When the model was applied to whole-slide images, the predominant subtype according to the artificial intelligence model classification matched that determined by pathologists in 75.5% of cases. The predominant subtype and tumor grade (using the WHO fourth and fifth classifications) determined by the artificial intelligence model resulted in similar recurrence-free survival curves to those determined by pathologists. Furthermore, we stratified the recurrence-free survival curves for patients with different proportions of high-grade components (solid, micropapillary and cribriform) according to the physical distribution of the high-grade component. The results suggested that tumors with centrally located high-grade components had a higher malignant potential (P < 0.001 for 5-20% high-grade component).

CONCLUSION

The new artificial intelligence model for histological subtyping of lung adenocarcinoma achieved high accuracy, and subtype quantification and subtype distribution analyses could be achieved. Artificial intelligence model therefore has potential for clinical application for both quantification and spatial analysis.

Artificial Intelligence and Lung Pathology

This manuscript provides a comprehensive overview of the application of artificial intelligence (AI) in lung pathology, particularly in the diagnosis of lung cancer. It discusses various AI models designed to support pathologists and clinicians. AI models supporting pathologists are to standardize diagnosis, score PD-L1 status, supporting tumor cellularity count, and indicating explainability for pathologic judgements. Several models predict outcomes beyond pathologic diagnosis and predict clinical outcomes like patients' survival and molecular alterations. The manuscript emphasizes the potential of AI to enhance accuracy and efficiency in pathology, while also addressing the challenges and future directions for integrating AI into clinical practice.

Ensemble Distillation of Divergent Opinions for Robust Pathological Image Classification

The construction of highly accurate deep neural networks (DNNs) requires consistent labeled data. However, there are numerous cases wherein the ground truth is not uniquely determined, even for the same data, owing to different interpretations depending on observers' decision criteria. Studies on the definition of labels and the building of a DNN model under such circumstances are scarce. Thus, this study addresses this issue in the field of pathological image diagnosis, where opinions occasionally vary, even among medical experts. We propose a method for constructing DNN models that are more robust to inter-observer variability by exploiting the knowledge about the relationships among the data learned by multiple DNN models. Comparison experiments were conducted using multiple pathology datasets of independently labeled images by different pathologists for the same set of images. The proposed method exhibited good generalization capability and outperformed the classification accuracy of baseline models specific to certain decision criteria.

Reproducibility of Assessment of Lepidic (Noninvasive) Patterns in Lung Adenocarcinoma With Cytokeratin Immunostain Compared With Hematoxylin and Eosin and the Proposed New International Association for the Study of Lung Cancer (IASLC) Algorithm

Introduction

Lepidic growth is considered noninvasive in lung nonmucinous adenocarcinoma, whereas other patterns are invasive. Considerable interobserver variability in assessing "invasion" has been reported. We assessed the utility of cytokeratin 7 (CK7) stain and recently proposed International Association for the Study of Lung Cancer criteria to improve assessment of noninvasion in lung adenocarcinoma.

Methods

Four pathologists (two staff, two trainees) assessed 158 hematoxylin and eosin (HE)- and CK7-stained slides of 108 pT1N0-2 nonmucinous lung adenocarcinoma cases. Scoring took place in four rounds. First, sections were independently scored for percentage of noninvasive or probable noninvasive and invasive or probable invasive patterns. Second, after a consensus scoring algorithm for CK7 was formulated, the slides were rescored. Subsequent third-round scoring was conducted only on HE slides using the 2023 International Association for the Study of Lung Cancer proposed criteria, and fourth-round scoring on both HE and CK7 slides simultaneously. Intraclass correlation coefficient (ICC) was calculated for each round. Recurrence-free survival was assessed using Cox proportional hazards regression methods.

Results

In the first two rounds, interobserver concordance was consistently higher with CK7 (ICC range = 0.44-0.6) than HE (range = 0.24-0.49) scores. The IASLC proposed algorithm improved ICC of HE scores to 0.60 (95% confidence interval: 0.52-0.67), and round 4 HE and CK7 combined improved ICC to 0.75 (95% confidence interval: 0.70-0.80). Continuous measures of averaged noninvasive and probable noninvasive scores on HE were associated with improved recurrence-free survival (hazard ratio: 0.83-0.86).

Conclusions

CK7 staining consistently increased interobserver concordance in assessment of invasive versus noninvasive patterns than HE. Combining CK7 with the 2023 IASLC criteria for morphologic features of invasion may further improve the interobservers' concordance for the recognition of lepidic growth in nonmucinous lung adenocarcinoma.

Invasion and Grading of Pulmonary Non-Mucinous Adenocarcinoma

Lung adenocarcinoma staging and grading were recently updated to reflect the link between histologic growth patterns and outcomes. The lepidic growth pattern is regarded as "in-situ," whereas all other patterns are regarded as invasive, though with stratification. Solid, micropapillary, and complex glandular patterns are associated with worse prognosis than papillary and acinar patterns. These recent changes have improved prognostic stratification. However, multiple pitfalls exist in measuring invasive size and in classifying lung adenocarcinoma growth patterns. Awareness of these limitations and recommended practices will help the pathology community achieve consistent prognostic performance and potentially contribute to improved patient management.

Standardized Classification of Lung Adenocarcinoma Subtypes and Improvement of Grading Assessment Through Deep Learning

The histopathologic distinction of lung adenocarcinoma (LADC) subtypes is subject to high interobserver variability, which can compromise the optimal assessment of patient prognosis. Therefore, this study developed convolutional neural networks capable of distinguishing LADC subtypes and predicting disease-specific survival, according to the recently established LADC tumor grades. Consensus LADC histopathologic images were obtained from 17 expert pulmonary pathologists and one pathologist in training. Two deep learning models (AI-1 and AI-2) were trained to predict eight different LADC classes. Furthermore, the trained models were tested on an independent cohort of 133 patients. The models achieved high precision, recall, and F1 scores exceeding 0.90 for most of the LADC classes. Clear stratification of the three LADC grades was reached in predicting the disease-specific survival by the two models, with both Kaplan-Meier curves showing significance (P = 0.0017 and 0.0003). Moreover, both trained models showed high stability in the segmentation of each pair of predicted grades with low variation in the hazard ratio across 200 bootstrapped samples. These findings indicate that the trained convolutional neural networks improve the diagnostic accuracy of the pathologist and refine LADC grade assessment. Thus, the trained models are promising tools that may assist in the routine evaluation of LADC subtypes and grades in clinical practice.

Efficacy of Preoperative Biopsy in Predicting the Newly Proposed Histologic Grade of Resected Lung Adenocarcinoma

A novel histologic grading system for invasive lung adenocarcinomas (LUAD) has been newly proposed and adopted by the World Health Organization (WHO) classification. We aimed to evaluate the concordance of newly established grades between preoperative biopsy and surgically resected LUAD samples. Additionally, factors affecting the concordance rate and its prognostic impact were also analyzed. In this study, surgically resected specimens of 222 patients with invasive LUAD and their preoperative biopsies collected between January 2013 and December 2020 were used. We determined the histologic subtypes of preoperative biopsy and surgically resected specimens and classified them separately according to the novel WHO grading system. The overall concordance rate of the novel WHO grades between preoperative biopsy and surgically resected samples was 81.5%, which was higher than that of the predominant subtype. When stratified by grades, the concordance rate of grades 1 (well-differentiated, 84.2%) and 3 (poorly differentiated, 89.1%) was found to be superior compared to grade 2 (moderately differentiated, 66.2%). Overall, the concordance rate was not significantly different from biopsy characteristics, including the number of biopsy samples, biopsy sample size, and tumor area size. On the other hand, the concordance rate of grades 1 and 2 was significantly higher in tumors with smaller invasive diameters, and that of grade 3 was significantly higher in tumors with larger invasive diameters. Preoperative biopsy specimens can predict the novel WHO grades, especially grades 1 and 3 of surgically resected specimens, more accurately than the former grading system, regardless of preoperative biopsy or clinicopathologic characteristics.

Deep learning-based diagnosis of histopathological patterns for invasive non-mucinous lung adenocarcinoma using semantic segmentation

OBJECTIVES

The application of artificial intelligence (AI) to the field of pathology has facilitated the development of digital pathology, hence, making AI-assisted diagnosis possible. Due to the variety of lung cancers and the subjectivity of manual evaluation, invasive non-mucinous lung adenocarcinoma (ADC) is difficult to diagnose. We aim to offer a deep learning solution that automatically classifies invasive non-mucinous lung ADC histological subtypes.

DESIGN

For this investigation, 523 whole-slide images (WSIs) were obtained. We divided 376 of the WSIs at random for model training. According to WHO diagnostic criteria, six histological components of invasive non-mucinous lung ADC, comprising lepidic, papillary, acinar, solid, micropapillary and cribriform arrangements, were annotated at the pixel level and employed as the predicting target. We constructed the deep learning model using DeepLab v3, and used 27 WSIs for model validation and the remaining 120 WSIs for testing. The predictions were analysed by senior pathologists.

RESULTS

The model could accurately predict the predominant subtype and the majority of minor subtypes and has achieved good performance. Except for acinar, the area under the curve of the model was larger than 0.8 for all the subtypes. Meanwhile, the model was able to generate pathological reports. The NDCG scores were greater than 75%. Through the analysis of feature maps and incidents of model misdiagnosis, we discovered that the deep learning model was consistent with the thought process of pathologists and revealed better performance in recognising minor lesions.

CONCLUSIONS

The findings of the deep learning model for predicting the major and minor subtypes of invasive non-mucinous lung ADC are favourable. Its appearance and sensitivity to tiny lesions can be of great assistance to pathologists.