MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Robustness Analysis of CNNs Against Similarity Transformations in Volumetric CT Data

D Eschweiler , T Stehle*, T Brosch , H Schulz , Philips Reseach, Hamburg, Germany

Presentations

(Sunday, 7/29/2018) 1:00 PM - 1:55 PM

Room: Karl Dean Ballroom C

Purpose: CT image datasets expose differing data due to varying patient positioning and body sizes. This study assesses the robustness of convolutional neural networks (CNNs) against such variations by evaluating segmentation performance on transformed data.

Methods: Utilizing a foveal fully convolutional network architecture (Brosch et al.), a CNN is trained for organ segmentation (liver, spleen, left and right kidney). The network is trained and evaluated on 70 standard CT data cases with consistent patient positioning obtained for radiotherapy planning (RTP) in a 5-fold cross-validation. Models are applied to transformed variations of the associated test data, including either rotation around the x-, y- or z-axis (-20° to +20°) or spatial scaling (factor 0.7 to 1.3). The Dice Similarity Coefficient (DSC) is computed as quality metric for different levels of each transformation and finally averaged over all splits.

Results: Increasing rotation proves to gradually decrease DSC, showing symmetric results for both the negative and positive direction. Strength of effect is characteristic to each spatial axis and each organ, resulting in nearly constant DSCs for small rotations of ±5° around any of the volume axes and DSC degradations ranging from 0.009 to 0.074 for rotations of ±20°. Scaling experiments show distinct susceptibility against downscaling up to a DSC decrease of 0.209 and less sensitivity against upscaling with a maximum DSC decrease of 0.069. Nearly constant DSCs are observed for scaling factors ranging from 0.9 to 1.1.

Conclusion: All experiments show robust segmentations for rotation and scaling within certain ranges, which is crucial for the application of CNNs to medical image processing. However, exceeding these ranges leads to a considerable decrease of segmentation quality. For robust segmentation, a consistent patient positioning and utilization of training data covering different patient sizes is highly recommended.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email