(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: investigate strategies of improving the performance of deep-learning-based OAR segmentation for gynecological patients, which usually suffers from limited training data due to the less patients available, by incorporating data from prostate cancer patients and optimizing network structure.
Methods: employed U-Net as the basic structure of our segmentation framework, but replaced each convolutional building block with a residual block where the input is directly added to the output of two consecutive convolutional operations. These long- and short-range skipped connections greatly facilitate gradient back-propagation and allow the decoding pathway to directly integrate high resolution features from the encoding pathway. In order to further boost its representative capability, we embedded channel-wise attention mechanism into the model by re-calibrating channel response at each residual block. The direct application of this model to challenging segmentation task such as OAR contouring on gynecological patients will suffer from over-fitting due to limited training data. So, three strategies were investigated: 1) incorporating prostate cancer patients to enlarge training set; 2) adding a classification branch in the original network to differentiate gynecological patients from prostate cancer patients and thus to characterize the anatomical difference between them; 3) oversampling the gynecological cases to account for the sample imbalance between gynecological and prostate cancer patients.
Results: direct application of the original model on 32 gynecological cases resulted in a mean Dice-Similarity-Coefficient (DSC) of 0.770 on bladder and 0.454 on rectum under four-fold cross-validation. The proposed strategies consistently improved the segmentation performance with 80 prostate cancer patients, achieving a statistically significant improved DSC of 0.857 on bladder (p = 0.003) and DSC of 0.496 on rectum.
Conclusion: preliminary results demonstrate that by carefully designing network structure and learning strategies, prostate cancer cases can be used to improve the segmentation performance of gynecological patients despite their significant anatomical deviations.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
