Program Information
Transfer Learning Using Deep Convolutional Neural Networks for Rectum Toxicity Prediction in Cervical Cancer Radiotherapy
X Zhen1,3*, J Chen3 , Z Zhong2 , B Hrycushko1 , L Zhou3 , S Jiang1 , K Albuquerque1 , X Gu1 , (1) The University of Texas Southwestern Medical Center, Dallas, TX, (2) Wayne State University, Detroit, MI, (3)Southern Medical University, Guangzhou, Guangdong
Presentations
TU-FG-605-3 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: We employ the deep convolutional neural network (CNN) with transfer learning technique for rectum dose-toxicity relationship prediction in cervical cancer treated with combined high-dose-rate brachytherapy (BT) and external beam radiotherapy (EBRT).
Methods: A VGG-16 CNN that pre-trained on a large scale natural image database ImageNet was built as our rectum toxicity prediction model. To fine-tune the pre-trained VGG-16 CNN, fractional doses including EBRT of the training cohort were deformedly summed by a previously developed topography-preserved point-matching DIR (TOP-DIR) algorithm. The cumulative EBRT+BT doses on the rectum surface were unfolded to obtain the rectum surface dose maps (RSDMs) which were then used to fine-tune the pre-trained VGG-16 network. The gradient-weighted class activation maps (Grad-CAM) that highlight the discriminative regions on the RSDM were also generated along with the CNN prediction model for analysis. The advantage of the proposed model over using logistic regression (LR) on dose volume parameters, i.e. the D0.1cc D1cc and D2cc (most exposed 0.1, 1 and 2-cm3 volume), for toxicity prediction is validated.
Results: The proposed prediction model was validated by a leave-one-out method on 42 (12 toxicity and 30 non-toxicity) cervical cancer patients. With all layers fine-tuned in VGG-16, a satisfactory prediction performance was achieved with total accuracy of 88.1%, sensitivity of 75%, specificity of 93.3% and AUC of 0.96, when compared the moderate prediction capability by LG with accuracy 64.3%, sensitivity 66.7%, specificity 58.3% and AUC 0.61 using D0.1/1/2cc. The salient regions of the Grad-CAMs of the toxicity groups were found locating on the upper rectum regions which corresponded to the regions with small p-values in the RSDM p-value map.
Conclusion: The proposed transfer learning CNN-based rectum toxicity prediction model has promising applications in cervical cancer radiotherapy and can be easily extended to prostate, pelvic, and other disease sites.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by Varian Medical Systems Inc (#OTD-109235)
Contact Email: