Click here to

Session: Deep Learning Image Processing and Segmentation [Return to Session]

A Cascade Neural Network for Tumor Segmentation in Head & Neck Cancer

P Ferjancic1*, H Liu2, A Qasem3, Z Zhou4, (1) University of Chicago, Chicago, IL, (2) Felucca Artificial Intelligence, Lenexa, KS,(3) ,Warrensburg, MO, (4) University of Kansas Medical Center, Kansas City, KS


SU-H400-IePD-F6-2 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Accurately segmenting tumor plays an important role on diagnosis and treatment in head & neck (H&N) cancer. We aim to develop a cascade neural network (CaNN) model to achieve automatic tumor localization and segmentation through a cascade of YOLO / U-Net in CT images.

Methods: In this study, totally 48 patients with PET and CT images in H&N obtained from The Cancer Imaging Archive are used. CaNN consists of four stages. The first stage is pre-processing. To correctly map the dimensions of PET images to the corresponding CT image, PET preprocessing is essential. On the other hand, it is important to enhance the contrast among the organs in CT image. The second stage is object localization model and segmentation model training, The third stage is tumor localization and segmentation. starts by feeding the pre-processed PET image into the trained YOLOv4 model, it detects the probability and location of the tumor in PET by producing a bounding box and probability, this bounding box is used in an intermediary step to automatically crop the contrast enhanced CT image into a fixed dimension that matches the U-Net, in cases where more than one bounding box is generated, only the box with the highest probability was considered. The cropped CT image is then fed into the U-Net for segmentation. The final stage is post-processing. OTSU threshold is applied on the segmentation mask to eliminate false positives.

Results: The mean and standard deviation value of CaNN on sensitivity, specificity, and F-score are 0.89±0.09, 0.99±0.00, 0.84±0.04, respectively, while U-Net is 0.77±0.00, 0.99±0.01, 0.71±0.28.

Conclusion: A new CaNN for automatic tumor segmentation in H&N cancer was proposed in this study. The experimental results demonstrated that CaNN can localize tumor very well and obtain promising segmentation performance as well.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email