Convolutional Neural Network Based Fully Automatic Lesion Localization in Rectal Cancer
Y Zhang1*, L Shi2, X Sun2, S Jabbour1, N Yue1, K Nie1, (1) Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, (2) Department of Radiation Oncology, Sir Run Run Shaw Hospital, Zhejiang Univ., Hangzhou, China
Presentations
SU-IePD-TRACK 1-1 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]
Purpose: To implement deep learning methodology involved in multi-modalities MRI imaging, to fully automatically localize and segment rectal tumors.
Methods: A total of 54 patients with stage T3 and T4 colorectal cancer were studied. The imaging protocol included T2w, T1w DCE and DWI. All images were reviewed by a single experienced oncologist specialized in rectal MRI. Tumors were outlined as a region of interest (ROI) on the third-phase of DCE. As input to the neural network architecture, we choose image patches with size of 21x21 pixels which locate at the same position on one slice from all frames. Using these patches, a 7 layer Convolutional Neural Network(CNN) was established to test whether the middle of the patch belong to the ROI. The output can be interpreted as the probability that the central pixel in the input image patch is associated with a tumor. Five-fold cross-validation was used to evaluate the performance of the CNN classifier. For the training process, a large number of batches were employed and extracted from different areas inside and outside ROIs. For the validation set, all of the image pixels were tested in the trained network architecture to determine whether each pixel belongs to the ROI areas.
Results: To evaluate the performance of the segmentation, the Dice Similarity Coefficient (DSC) was used to compare the results of the proposed algorithm and the ROI outlined by reader. Among all 54 patients, the mean of DSC is 0.76.
Conclusion: Based on the results, CNN can be utilized to develop a rectal cancer segmentation method. This can provide helpful information for rectal cancer patients to assess their risk more accurately to choose an optimal strategy when designing radiation therapy planning.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email