Virtual Library
Start Your Search
Hugo Aerts
Author of
-
+
MA18 - Modelling, Decision-Making and Population-Based Outcomes (ID 920)
- Event: WCLC 2018
- Type: Mini Oral Abstract Session
- Track: Treatment in the Real World - Support, Survivorship, Systems Research
- Presentations: 1
- Moderators:
- Coordinates: 9/25/2018, 13:30 - 15:00, Room 201 F
-
+
MA18.01 - Non-Small Cell Lung Cancer Risk Assessment with Artificial Neural Networks (ID 13532)
13:30 - 13:35 | Author(s): Hugo Aerts
- Abstract
- Presentation
Background
Lung cancer is a heterogeneous disease with many clinically important subtypes. Given the complexity of classification, there is room for innovative risk assessment tools to help ascertain prognosis and management. In this work we tested an Artificial Neural Network (ANN) to stratify patients into clinically significant low and high risk categories.
a9ded1e5ce5d75814730bb4caaf49419 Method
CT imaging, survival, and cancer staging data was extracted for a sample of 311 patients with Stage-I (n = 186) and Stage-II (n = 125) non-small cell lung cancer (NSCLC) from the comprehensive Boston Lung Cancer Survival (BLCS) cohort. Median follow-up from time of diagnosis was 3.5 years, with 86% 2-year survival. A deep convolutional neural network pretrained on ImageNet was used, with fine-tuning of the last convolutional layers, dense layers, and softmax for stratification. Inputs of this model were 50 x 50 mm2 image patches. Training was performed on 182 labeled CT scans (112 Stage-I and 70 Stage-II). 46 cases were used for initial cross-validation, with an independent test set of 83 cases. The median prediction probability from the ANN was used as a cutoff to divide patients into low and high risk groups.
4c3880bb027f159e801041b1021e88e8 Result
The model was able to perform classification of cancer stage on the heterogeneous test set (AUC = 0.73, p< 0.0005). The test set was split evenly into low risk (n = 42) and high risk (n= 41) groups based on model predictions. There was statistically significant separation in the Kaplan Meier-estimates for survivorship in the two stratified groups (p < 0.02).
8eea62084ca7e541d918e823422bd82e Conclusion
ANNs can be effective tools for quantitative risk stratification in NSCLC. In addition to the potential for real-time clinical decision support, ANNs may also help create new paradigms in lung cancer risk assessment. The models have the capacity to perform suprahuman computations, which can help meet future demands of clinical practice, given expanding digital-imaging volumes.
6f8b794f3246b0c1e1780bb4d4d5dc53Only Members that have purchased this event or have registered via an access code will be able to view this content. To view this presentation, please login, select "Add to Cart" and proceed to checkout. If you would like to become a member of IASLC, please click here.