L Chen*, K Wang, C Shen, D Sher, J Wang, UT Southwestern Medical Center, Dallas, TX

• 
  L Chen

• 
  

WE-F-TRACK 6-1 (Wednesday, 7/28/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: Accurate estimation of the risk of loco-regional recurrence (LRR) for patients with head and neck squamous cell cancer (HNSCC) prior to treatment is essential to physicians for making personalized treatment management. In this work, we proposed a bilateral neural network to predict LRR through using pre-treatment CT, PET and clinical data.

Methods: In the proposed scheme, there are two classification convolutional neural network (cCNN) modules: the first cCNN module (cCNN_1) uses region of interest (ROI) containing tumor and its surrounding tissues as input while the second cCNN module (cCNN_2) uses tumor itself as input. cCNN_1 pays attention on the information of the tumor position and tumor contrast compared to surrounding tissues, while cCNN_2 concerns more on the tumor inside and tumor boundary. The features learned by these two networks are complementary, which are then fused with clinical data for the final LRR prediction. Additionally, due to the highly-imbalanced characteristics of the dataset (with vs. without LRR), we proposed a new loss function termed as focal-dual loss which focuses more on the samples hard to classify and also tries to decrease the probabilities that a data point is assigned to an incorrect class to train the classification model. This study included the images and clinical variables of 316 patients with HNSCC received radiation treatment, among which 60 patients experienced LRR. Five folder cross validation was used to evaluate the performance of the proposed scheme.

Results: The sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) values obtained by the proposed scheme were 0.77, 0.75, 0.75 and 0.80 respectively, outperforming conventional cCNNs only using ROI or tumor itself, and radiomics approaches.

Conclusion: The proposed method combining clinical feature and CT & PET imaging features learned from two cCNN modules provides a more accurate way to predict HNSCC LRR.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

IM/TH- Informatics: Informatics in Therapy (general)

Contact Email