MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Self-Attention Based Deep Learning Probabilistic Parotid Gland Segmentation Quality Evaluation Using Dose Volume Histogram Analysis

S Berry*, J Jiang , S Elguindi , M Hunt , J Deasy , H Veeraraghavan , Memorial Sloan Kettering Cancer Center, New York, NY

Presentations

(Tuesday, 7/16/2019) 3:45 PM - 4:15 PM

Room: Exhibit Hall | Forum 5

Purpose: To compare the quality of probabilistic segmentations using self-attention based Unet (UnetSA) to a basic Unet (Unetbasic) by observing deviations in dose volume histograms (DVH) in addition to typical geometry based structure quality metrics (SQMgeo).

Methods: We developed two deep learning networks (UnetSA, Unetbasic) for segmentation trained on 48 head-and-neck cancer (HN) patient CTs. Both networks output probabilistic segmentations using soft-max function, which pools the activations from various features to produce votes for individual pixels as belonging to the structure or background. Using this map, multiple parotid segmentations were computed using thresholds of 0.1, 0.3, 0.5, 0.7 and 0.9 and DVH curves were constructed from those segmentations using the dose distribution planned from the expert manual segmentations (EMS). The trained models were evaluated on an independent set of 22 patients not used in training. The segmentation accuracy was evaluated using the 0.1 segmentation threshold with the Dice Similarity Coefficient (DSC) and 95th percentile of the Hausdorff Distance (HD95). The width of the DVH band resulting from the segmentations produced at thresholds of 0.1 and 0.9 (DVHwidth), and the difference in relative volume receiving 26Gy (ΔV26Gy) and 30Gy (ΔV30Gy) between the center of that band and the EMS was calculated.

Results: The UnetSA produced more accurate segmentation for left (DSC 0.84±0.04; HD95 2.23±0.06) and right (DSC 0.85±0.04; HD95 2.34±0.073) parotid glands compared with Unetbasic (Left parotid DSC 0.83±0.03, HD95 2.47±0.54; right parotid DSC 0.84±0.04, HD95 2.21±0.54), respectively. The DVHwidth at 26Gy using the best method was 0.026±0.014 and right parotid ΔV26Gy was 0.038±0.037.

Conclusion: UnetSA led to more accurate segmentations and more closely approximated the DVH from the EMS. In addition to improving a basic Unet architecture, we also studied how incorporating DVH metrics can frame judgments of segmentation quality in terms of relevance to radiotherapy planning.

Funding Support, Disclosures, and Conflict of Interest: This work was partially funded by a Master Research Agreement with Varian Medical Systems and was also partially funded by NCI R01 CA198121.

Keywords

Segmentation, Radiation Therapy, Dose Volume Histograms

Taxonomy

IM/TH- image segmentation: CT

Contact Email