Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Using New Feature Extraction Framework to Predict Heart Toxicity in Breast Cancer

B Choi1*, S Yoo1,J Moon1, J Kim1, J Chang1, H Kim1, (1) Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul

Presentations

PO-GePV-M-263 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Radiation-induced cardiac toxicity such as acute coronary event (ACE) is a key issue of breast radiotherapy. As it has been mostly discussed in mean heart dose, however, there are few studies about cardiac substructures and ACE after radiotherapy. This work proposes a new heart toxicity prediction model by a new feature extraction and processing framework with cardiac sub-structure and dose distribution.

Methods: The patient cohort consists of 29 toxicity and 16 non-toxicity cases after breast radiotherapy. A recent study demonstrated that the feature extraction from segmenting target volume followed by feature selection produced better prediction accuracy in survival prediction after lung SBRT than radiomics and conventional deep neural network (convolution/fully-connected layers). We applied the framework for toxicity prediction. It compared three types of prediction models: a) radiomics, b) conventional deep neural network, and c) feature extraction and processing model with CT image and dose information as input, and heart segmentation as output. The output of the auto-segmentation model in (c) was varied from entire heart to 6 sub-structures and 3 sub-structures (left-ventricle, right-ventricle, and left-anterior descending artery). The relationship between the features and prediction was visualized by a method, called Grad-CAM. The network training and testing was conducted with 4-fold cross-validation.

Results: The conventional deep neural network (50-60% accuracy) performed worst. The proposed model led to higher prediction accuracy (over 78%) than radiomics (69%). Impressively, the predictive model with 6 and 3 sub-structures (81%) outperformed that with entire heart structure (78%). The model with 3 sub-structures showed greater specificity and sensitivity than that with 6-substructures, though both having the same accuracy. When visualized by Grad-CAM, the hot-spots were generated around left-ventricle region in most cases.

Conclusion: Our proposed prediction model characterized by 1) new feature extraction/processing process and 2) cardiac sub-structure dose/structure information enhanced the prediction accuracy.

ePosters

    Keywords

    Segmentation

    Taxonomy

    IM- Radiation Dose and Risk: General (Most Aspects)

    Contact Email

    Share: