TL;DR
This paper introduces two deep learning models (AI-1 and AI-2) that can accurately distinguish lung adenocarcinoma (LADC) subtypes and predict patient survival based on histopathological images. These AI models demonstrate performance comparable to, and in some cases superior to, expert pathologists, offering a promising tool to standardize LADC diagnosis and improve patient prognostication.
What Problem Does This Paper Address?
Diagnosing and grading lung adenocarcinoma (LADC) relies heavily on histopathological examination of tissue samples. However, distinguishing between the various LADC subtypes and assigning a grade can be subjective, leading to significant inter-observer variability among pathologists. This variability can compromise the accuracy of patient prognosis and treatment decisions. The paper aims to address this challenge by developing automated deep learning models that can consistently and accurately perform these critical diagnostic tasks.
Methods
The researchers developed two deep learning models, AI-1 and AI-2, using the EfficientNet-B3 architecture. To create a reliable "ground truth" for training, they employed a clustering approach on consensus labels from 17 expert pulmonary pathologists and one trainee. This resulted in two distinct datasets, which were used to train AI-1 and AI-2, respectively. The models were trained to classify eight LADC categories: lepidic, acinar, papillary, micropapillary, solid, invasive mucinous adenocarcinoma (IMA), other carcinoma types, and no carcinoma cells. Data augmentation techniques, including pixel-level and spatial-level transforms and the mixup principle, were used to enhance the training data. Model validation was performed using k-fold cross-validation, and performance was evaluated using precision, recall, F1-scores, and Receiver Operating Characteristic (ROC) curves. The models' ability to predict disease-specific survival (DSS) based on LADC grades was also assessed and compared to pathologist predictions.
Datasets
- Training Data: Consensus LADC histopathological images derived from expert pathologist annotations, split into two datasets for AI-1 and AI-2 training based on a clustering approach.
- Testing Data: An independent cohort of 133 patients with LADC, with case-level diagnoses provided by 15 pulmonary pathologists.
Key Insights and Ideas
- AI as a Standardizing Force: The core idea is that AI can overcome the inherent subjectivity in histopathology, providing a consistent and reproducible diagnostic tool.
- Clustering for Ground Truth: Using a clustering approach to group pathologists with similar interpretations was a novel way to establish a more robust and less biased ground truth for training.
- Survival Prediction from Histology: The study goes beyond simple classification to predict patient survival, directly linking microscopic features to clinical outcomes, which is a significant step for AI in pathology.
- AI Outperforming Humans (in some aspects): The finding that one of the AI models was more accurate than 14 out of 15 pathologists in grading is a powerful demonstration of AI's potential.
Key Results
- High Classification Accuracy: Both AI-1 and AI-2 achieved high performance in predicting LADC classes, with accuracies exceeding 97% and 95%, respectively.
- Excellent Precision and Recall: For most LADC classes, precision, recall, and F1-scores exceeded 0.90. The "no carcinoma," "IMA," and "solid" classes showed particularly high precision for AI-1, while AI-2 performed similarly for "no carcinoma" and "solid."
- Effective Survival Stratification: The trained models demonstrated clear stratification of disease-specific survival (DSS) across the three LADC grades, with a reduction in DSS as grades increased.
- Superior Grading Performance: AI-2 showed better statistical significance in DSS prediction compared to AI-1 and outperformed 14 out of 15 pathologists (p=0.0003).
- Stable Performance: Both models exhibited low variation in hazard ratio predictions across bootstrapped samples, indicating stable performance in grading.
Limitations and Future Work
- Single-Institutional Data: The study used cases from a single institution and scanned with a single scanner, which might limit the generalizability of the models.
- Limited "Other Carcinoma Types" Data: The "other carcinoma types" category had fewer training tiles and lower agreement among pathologists, impacting the models' precision for this subtype.
- One Cohort Evaluation: The models were evaluated on only one cohort, and further evaluation on multiple cohorts from different institutions is recommended to ensure widespread applicability.
- Future Work: The authors suggest increasing the training set size and including more complex glandular tumor growth or rare LADC subtypes to refine prediction accuracy. They also emphasize the need for multi-institutional validation to generalize the models for broader clinical use.
Summary updated