MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automatic Segmentation of the Prostate On CT Images Using a Bi-Directional Convolutional LSTM U-Net with Novel Loss Function

X Li1*, H Bagher-Ebadian2, C Li1, E Mohamed2, F Siddiqui2, B Movsas2, D Zhu1, I Chetty2, (1) Wayne State University (2) Henry Ford Health System, Detroit, MI

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Deep learning-based networks have achieved practical performance in the field of medical image segmentation. Here we introduce a deep neural network with an enhanced U-net architecture trained using a novel loss function to perform automatic segmentation of the prostate gland on pre-treatment planning CT images of patients with prostate cancer.

Methods: Planning-CT datasets for 100 prostate cancer patients were retrospectively evaluated as part of an IRB-approved protocol. Physician-delineated images constituted the ground-truth. All images were linearly interpolated/resampled to a spatial resolution of 1×1×1.5mm. Image regions containing the cross-section of the prostate gland (128×128mm2) were used for training. Patient dataset was split into cohorts of train (55%), validation(15%), and test (30%) using a bi-directional convolutional long short-term memory (ConvLSTM) U-Net with densely connected convolutions (BCDU-Net).Two region-based loss functions (Focal Tversky Loss, FTL and Dice Loss, DL) were weighted combined to minimize the mismatch and maximize the overlapping regions between the ground-truth and the predicted segmentation. The FTL function is especially useful in dealing with extreme foreground-background class imbalance, as is the case with segmentation of the prostate gland vs. the surrounding background.

Results: Results comparing the optimal BCDU-Net (FTL+DL) and physician-generated contours for the test data were: DSC (Dice) =0.84, Area-under-receiver-operator-characteristic-curve (AUC=0.98), Area-under-precision-recall-curve (AUPRC=0.98), JSC (Jaccard) =0.98, Sensitivity=0.82, Specificity=0.99, Accuracy=0.99, and Precision=0.85. Results showed improved performance when the combined loss function vs. dice-based loss function only were used for training the deep neural network (see supporting doc).

Conclusion: Despite the small sample size BCDU-Net-based model utilizing a weighted loss function combining dice and focal tversky losses showed promise for automatic segmentation of the prostate gland on planning CT image datasets. Further investigation is warranted.

Download ePoster [PDF]

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a grant from Varian Medical Systems (Palo Alto, CA)

Keywords

Segmentation, CT, Prostate Therapy

Taxonomy

IM/TH- image Segmentation: CT

Contact Email