Name:
A Deep Dive into the RAISE Project Methodology
Description:
A Deep Dive into the RAISE Project Methodology
Thumbnail URL:
https://cadmoremediastorage.blob.core.windows.net/d8b28f68-25c3-4cc7-b2ba-6feec0090cb5/videoscrubberimages/Scrubber_5.jpg
Duration:
T00H08M35S
Embed URL:
https://stream.cadmore.media/player/d8b28f68-25c3-4cc7-b2ba-6feec0090cb5
Content URL:
https://cadmoreoriginalmedia.blob.core.windows.net/d8b28f68-25c3-4cc7-b2ba-6feec0090cb5/VJBM-2023-0018 RAISE Video v0.2.mp4?sv=2019-02-02&sr=c&sig=QsRTUhLh2%2FQR4INGiKE3YW%2F9Env%2FRpyFXZrEqk9ls4g%3D&st=2025-02-05T22%3A25%3A37Z&se=2025-02-06T00%3A30%3A37Z&sp=r
Upload Date:
2025-02-05T00:00:00.0000000
Transcript:
Language: EN.
Segment:1 Introduction.
SPEAKER: This animation discusses the methodology used in the Research for Artificial Intelligence -based Surrogate Endpoint project called RAISE. This exploratory work is presented in two manuscripts published in Future Oncology.
Segment:2 Background.
SPEAKER: Neuroendocrine tumors are rare, slow-growing tumors. These tumors can develop in many organs, but they often spread to the liver. Response Evaluation Criteria in Solid Tumors, or RECIST, is commonly used to assess treatment response in patients with neuroendocrine tumors in clinical trials. RECIST uses the sum of the longest diameter of lesions to classify a tumor response as either a complete response, partial response, progressive disease, or stable disease. In this way, RECIST is used to estimate the change in tumor burden over a course of treatment and to define progression-free survival, which is often used as a primary endpoint in clinical trials in patients with neuroendocrine tumors.
SPEAKER: However, as neuroendocrine tumors often grow slowly, it is difficult to tell whether a treatment has any effect by assessing the growth and shrinkage of these tumors using RECIST. A better way of measuring how well neuroendocrine tumors respond to treatment is needed to ensure that patients receive the right treatment as early as possible. The RAISE project aimed to create a new endpoint using a type of artificial intelligence called deep learning. This endpoint would allow earlier prediction of treatment response and progression-free survival in patients with neuroendocrine tumors.
Segment:3 The RAISE Project - Deep Learning Methodology .
SPEAKER: In RAISE, we analyzed CT scans from a subset of patients in the CLARINET phase three trial. Patients with liver lesions and CT scans at multiple time points were selected for the deep learning analysis to explore the potential of deep learning models in predicting treatment response. Patients with multiple lesions were selected for the analysis of response heterogeneity to assess the value of response heterogeneity as a biomarker for earlier prediction of progression-free survival.
SPEAKER: Deep learning features were extracted from images using a convolutional neural network. The convolutional network in this study was trained on data from ImageNet, a large dataset of 3.2 million images. Features were extracted via transfer learning methodology. This methodology uses features learned for a different task on a larger dataset and assumes that these features are relevant enough to be applied to a smaller dataset of interest. In RAISE, the larger dataset was ImageNet, and the smaller dataset was imaging data from CLARINET.
SPEAKER: To select deep learning components, we used principal component analysis. This is a technique that aims to extract information and represent this extracted information as a set of new variables called principal components. Principal component analysis was used to reduce more than 2,000 deep learning features to 20 deep learning components. Then the features were aggregated to predict progression-free survival. To achieve this, the 20 selected components were analyzed via a feedforward artificial neural network called a multi-layer perceptron, which identified imaging features associated with progression.
SPEAKER: Different models were then used to assess the features captured by deep learning. The first model included a binary mask, which represents the contour of the lesion versus a lesion mask, which represents the entirety of the lesion. The second model assessed the deep learning features captured when using lesion-only versus whole-liver image inputs. The third model compared features captured when using lesion images obtained in the portal versus arterial contrast enhancement phase.
SPEAKER: A three-variable Cox model and a two-variable Cox model were then used to evaluate how well deep learning models could predict progression-free survival compared with currently available markers of progression, such as the sum of the longest diameter ratio and levels of chromogranin A, measured in the blood.
Segment:4 The RAISE Project - Response Heterogeneity Methodology .
SPEAKER: As part of the RAISE project, we also assessed whether response heterogeneity could be used as a biomarker to allow earlier prediction of progression-free survival in patients with neuroendocrine tumors. Response heterogeneity has been defined previously based on RECIST as the situation in which some tumors respond well to treatment while other tumors in the same patient do not. However, the RECIST-based definition of response heterogeneity is not suitable for slow-growing tumors, such as neuroendocrine tumors. So this definition was adapted in RAISE.
SPEAKER: To define response heterogeneity in RAISE, we measured the longest diameter of target lesions on CT images at baseline and subsequent patient visits. The ratio between the longest diameter of target lesions at each visit and baseline was then calculated. Response heterogeneity was estimated using the standard deviation of the longest diameter ratio of all target lesions for each patient. Next, the value of response heterogeneity in predicting progression-free survival was evaluated using Cox models.
Segment:5 Results.
SPEAKER: For the deep learning analysis, the lesion mask model demonstrated improved performance compared to the binary mask model at week 72. This finding shows that deep learning models can be trained to capture additional information from images that may be relevant to the prediction of progression-free survival. However, no differences were found in the performance of a model combining the deep learning model output with the sum of the longest diameter ratio and levels of chromogranin A as compared with a model based on the sum of the longest diameter ratio and levels of chromogranin A alone. This finding showed that deep learning models could not improve the prediction of progression-free survival compared with other currently used biomarkers.
SPEAKER: For the response heterogeneity analysis, the reported hazard ratio for response heterogeneity at week 12 was 1.48, which was greater by week 36 at 1.72. The hazard ratio with control for the sum of the longest diameter ratio was 1.28 at week 12, with a greater value of 1.81 reported at week 36. These findings demonstrated that greater response heterogeneity was associated with faster progression, independent of the sum of the longest diameter ratio.
Segment:6 Conclusions.
SPEAKER: Overall, the RAISE project found that deep learning models can detect features in images of neuroendocrine tumors other than tumor size and shape, which may be relevant to the prediction of progression-free survival. However, deep learning models did not improve the prediction of progression-free survival in patients with neuroendocrine tumors compared with other currently used biomarkers. It was also found that response heterogeneity may be a predictor of progression-free survival in patients with neuroendocrine tumors, regardless of other currently available markers of progression. Further validation of this biomarker in a larger group of patients is required. [MUSIC PLAYING]
Segment:7 Closing.
SPEAKER: [MUSIC PLAYING]
Form Title Download
Form Title Bookmark
Form Title Share
Form Title Cite
