Deep learning-based survival prediction for multiple cancer types using histopathology images (1912.07354v1)

Abstract: Prognostic information at diagnosis has important implications for cancer treatment and monitoring. Although cancer staging, histopathological assessment, molecular features, and clinical variables can provide useful prognostic insights, improving risk stratification remains an active research area. We developed a deep learning system (DLS) to predict disease specific survival across 10 cancer types from The Cancer Genome Atlas (TCGA). We used a weakly-supervised approach without pixel-level annotations, and tested three different survival loss functions. The DLS was developed using 9,086 slides from 3,664 cases and evaluated using 3,009 slides from 1,216 cases. In multivariable Cox regression analysis of the combined cohort including all 10 cancers, the DLS was significantly associated with disease specific survival (hazard ratio of 1.58, 95% CI 1.28-1.70, p<0.0001) after adjusting for cancer type, stage, age, and sex. In a per-cancer adjusted subanalysis, the DLS remained a significant predictor of survival in 5 of 10 cancer types. Compared to a baseline model including stage, age, and sex, the c-index of the model demonstrated an absolute 3.7% improvement (95% CI 1.0-6.5) in the combined cohort. Additionally, our models stratified patients within individual cancer stages, particularly stage II (p=0.025) and stage III (p<0.001). By developing and evaluating prognostic models across multiple cancer types, this work represents one of the most comprehensive studies exploring the direct prediction of clinical outcomes using deep learning and histopathology images. Our analysis demonstrates the potential for this approach to provide prognostic information in multiple cancer types, and even within specific pathologic stages. However, given the relatively small number of clinical events, we observed wide confidence intervals, suggesting that future work will benefit from larger datasets.

• The paper introduces a CNN-based deep learning system that predicts survival in 10 cancer types without laborious pixel-level annotations.
• It achieves notable prognostic value with a hazard ratio of 1.58 and a 3.7% improvement in the c-index compared to traditional models.
• The research demonstrates the potential of weakly-supervised deep learning to discover novel morphologic features for personalized oncology treatments.

Deep Learning-Based Survival Prediction for Multiple Cancer Types Using Histopathology Images

The paper presents a comprehensive paper on leveraging deep learning methodologies to predict disease-specific survival across multiple cancer types using histopathology images sourced from The Cancer Genome Atlas (TCGA). The proposed deep learning system (DLS) emphasizes a weakly-supervised approach, thereby eliminating the need for pixel-level annotations which are typically laborious and time-consuming to acquire.

Methodological Framework

The research employs a convolutional neural network (CNN) architecture, leveraging depth-wise separable convolution layers similar to MobileNet. The CNNs analyzed image patches extracted from whole-slide histopathology images (WSIs) and processed them through an average pooling layer and a final logistic regression layer to generate survival predictions. The scope of the paper included 10 diverse cancer types, and the research identified three distinct loss functions to handle the non-trivial task of survival prediction under right-censored conditions. The chosen architecture and training techniques underscore the research's intention to handle large image sizes and morphological heterogeneity in images without specific expert-annotated features.

Key Results

The DLS demonstrated a significant association with disease-specific survival in a multivariable Cox regression analysis. The system achieved a hazard ratio of 1.58 in a combined analysis of all cancer types, indicating its potential as a significant prognostic tool even after adjustments for variables like cancer type, stage, age, and sex. Notably, the DLS consistently outperformed baseline models that included traditional factors of age, sex, and tumor stage, witnessing an improvement of 3.7% in the c-index for the combined cohort.

Implications and Future Prospects

This work has significant implications for oncology, primarily as it furthers the potential utility of deep learning models in enhancing prognostic predictions without exhaustive expert annotations. By leveraging histopathology images, the paper suggests that novel morphological features of prognostic relevance can be discovered. Such insights could lead the way toward more refined stratification of patients, permitting personalized treatment regimens and monitoring strategies.

However, the paper faces limitations associated with small datasets relative to typical deep learning tasks, leading to wide confidence intervals that make effect sizes difficult to ascertain. Future developments could benefit from larger, more diverse datasets that more accurately reflect real-world clinical settings, with an exploration into incorporating additional variables such as molecular data for improved prediction accuracy.

Conclusion

The paper provides valuable insights into the potential of deep learning in cancer prognosis, demonstrating that significant predictive information can be gleaned from histopathology images in a weakly-supervised framework. By identifying survival-relevant morphologic features, the research presents an encouraging paradigm for augmenting existing clinical decision-making processes. Future work will undoubtedly focus on expanding datasets and refining algorithms to bolster the predictive power and clinical utility of such systems.