Fully Automated Organ Segmentation in Male Pelvic CT Images

INTRODUCTION

• Low prostate-to-background contrast and patient variation in organ shape, size, and intensity, together with diverse CT acquisition conditions, make automated pelvic-organ segmentation challenging.Variation includes contrast-agent effects, differing fields of view, and implanted fiducial markers or catheters.Source: INTRODUCTION · pp. 1–2
• Prior shape-model and machine-learning methods can rely on same-patient prior-treatment samples and hand-crafted features, while a gap remains between automated results and manual annotations.Source: INTRODUCTION · p. 2
• U-Net uses skip connections between feature maps at matching depths to preserve local features while propagating global features to higher-resolution layers.Source: INTRODUCTION · p. 2
• CT segmentation is desirable for radiotherapy planning because current workflows use CT for contouring and dose calculation, although MRI has better soft-tissue contrast for delineating boundaries.Source: INTRODUCTION · p. 2
• The paper presents a fully automated pelvic-CT multi-organ workflow combining 2D U-Net localization with 3D U-Net segmentation using ResNeXt blocks.The authors describe it as a deep-learning-only workflow and claim improved accuracy over existing methods.Source: INTRODUCTION · p. 3

METHODS

• Organ segmentation comprises localization of an organ’s whereabouts and delineation of its spatial extent and composition.Source: METHODS · p. 3
• Because only about 15% of pelvic CT slices contain prostate and localization errors can mislead segmentation, the workflow adds a 2D U-Net to detect organs automatically before segmentation.Source: METHODS · p. 3

Automation workflow

• The workflow crops the upper and lower 15% of images to reduce memory use; the cropping margin was determined for UT Southwestern patients and may be removed without affecting accuracy.Source: Automation workflow · p. 3
• The CT volume is downsized for speed and memory, localized organ volumes are extracted, and those volumes are passed to the final segmentation network.Source: Automation workflow · p. 3, Figure 1 · p. 4

Localization network

• The five-channel 2D U-Net assigns one channel to each organ and takes the corresponding CT image and supervised-training binary mask as channel inputs.It predicts five organ feature maps and is trained with a custom weighted Dice loss using Adam.Source: Localization network · p. 4, Figure 2 · p. 5

Segmentation network

• The 3D U-Net uses ResNeXt blocks in its encoding arm because 3D kernels can capture organ shape and surface across sequential slices, unlike 2D curves alone.Source: Segmentation network · p. 5, Figure 3 · p. 6
• Each ResNeXt block aggregates n paths of two 3×3×3 convolutions, concatenates their outputs, and combines the result with the block input before encoding-arm max pooling.Source: Segmentation network · p. 5
• The segmentation network receives localized organ volumes with fixed slice counts—32 for prostate and 64 for each organ at risk—and is trained with boundary-weighted Dice loss using Adam.Source: Segmentation network · p. 6

Loss functions

• The Dice similarity coefficient measures agreement between two image regions and is widely used to evaluate medical-volume segmentation against ground truth.Source: Loss functions · p. 6
• DSC is useful as a segmentation-network loss because it efficiently addresses class imbalance.Source: Loss functions · p. 6
• The ground-truth and predicted foreground-voxel counts are the quantities denoted by the notation in the DSC definition.Source: Loss functions · p. 6

Organ weighted dice function for localization network

• For localization, annotated pixels receive weights inversely proportional to their class sample counts, so organs contribute more equally to the Dice loss.The organ-specific weight vector is determined during training from the number of samples belonging to that organ.Source: Organ weighted dice function for… · p. 7

Boundary weighted dice function for segmentation network

• For segmentation, boundary-weighted Dice assigns greater importance to pixels nearer the organ boundary, increasing their contribution to the loss.The border weight matrix is determined during training.Source: Boundary weighted dice function for… · p. 7

• The dataset included raw CT scans from 136 prostate cancer patients collected at the University of Texas Southwestern Medical Center.Source: EXPERIMENTS & RESULTS Patients Data · p. 7
• Images were acquired using a 16-slice CT scanner with a 512x512 matrix and 2 mm slice thickness.The reported voxel size was 1.17mm×1.17mm×2mm.Source: EXPERIMENTS & RESULTS Patients Data · p. 7
• Experienced radiation oncologists contoured the prostate, bladder, rectum, and femoral heads.Source: EXPERIMENTS & RESULTS Patients Data · p. 7

Training Data

• Four-fold leave-one-out cross-validation reserved one group for validation, and the four models’ predictions were averaged on the test set.Eighty percent of patient data were used for training and the remaining 20% for testing.Source: Training Data · p. 7, Figure 4 · p. 8
• ResNeXt blocks were placed along the 3D U-Net encoding path to gain accuracy through cardinality without making the network too deep or computationally infeasible.The stated hardware constraint was an NVIDIA Tesla K80 dual-GPU card.Source: Training strategies · p. 8
• The model settled on 3×3×3 convolution kernels after 5×5×5 kernels produced inaccurate edges and small contours; a 1e-2 learning rate with plateau-based decay gave the best validation Dice.Source: Training strategies · p. 8
• Early stopping and dropout addressed overfitting, batch normalization was added for stability, and learning rate was decayed when validation loss plateaued.Source: Training strategies · p. 8
• Deep-learning models were implemented in Keras and processed on an NVIDIA Tesla K80 dual-GPU card.Source: Training strategies · p. 8

Results

• Predicted segmentations closely resemble manual segmentations in the shown femoral-head, prostate, bladder, and rectum examples, including prostate base, middle, and apex slices.Figures 5 and 6 show organ examples, while Figure 7 shows prostate slices from base to apex.Source: Results · p. 9, Figure 5 · p. 9, Figure 6 · p. 9
• The averaged prostate model’s maximum DSC was 93% and minimum DSC was 85.9%.Figure 8 plots prostate DSC values for the four cross-validation models and the final averaged model.Source: Results · p. 10, Figure 8 · p. 10
• Cross-validation results for bladder, prostate, and rectum are summarized in Table 1, and Figure 9 shows Dice-coefficient distributions for all segmented organs.Source: Results · p. 10, Table 1 · p. 10, Figure 9 · p. 10
• The final test evaluation averaged predictions from four cross-validation models and reported average DSC, average surface distance, and average Hausdorff distance for DSC loss, plus average DSC for BCE loss.Table 2 also reports significance tests for the improvement when using weighted DSC loss.Source: Results · p. 11, Table 2 · p. 11

DISCUSSION

• Mean DSCs were 0.90 for prostate, 0.95 for bladder, 0.84 for rectum, 0.95 for the left femoral head, and 0.96 for the right femoral head.After training, test-set prediction took a few seconds; the full workflow took a few minutes per patient.Source: DISCUSSION · p. 11
• Rectum DSC was 0.84 with large variation, likely reflecting patient variation in organ size and shape; rectal-balloon cases were predicted well.The authors also suggest that reliance on MRI for many manual rectum segmentations may help explain the suboptimal results.Source: DISCUSSION · p. 11, Figure 10 · p. 12
• Direct comparison with published methods was not possible, although Table 3 compares reported DSC and surface-distance metrics.The surface measures are average surface distance and average Hausdorff distance.Source: Table 3 · p. 12, DISCUSSION · p. 12
• 90.2% cross-validation balanced accuracy for prostate and the lowest computation time marked the proposed 3D network as best among the tested U-Net variants.The architectures used the same patient dataset but were individually fine-tuned with different layer counts, depths, and dropout rates.Source: Table 4 · p. 13, DISCUSSION · p. 13
• All tested networks performed poorly without boundary-weighted Dice loss; Figure 11 contrasts predictions from weighted and normal Dice loss.The figure shows weighted-loss predictions in blue, normal-loss predictions in green, and ground truth in red.Source: Figure 11 · p. 13, DISCUSSION · p. 13
• Manual segmentation is an imperfect ground truth because of inter-observer variability; the authors propose averaging multiple observers' contours in future work.They also propose an ensemble network to combine organ segmentations without overlap.Source: DISCUSSION · p. 14

CONCLUSION

• The hybrid network uses CT images as its only input and requires no manual intervention for segmentation.The authors state that the workflow could be adapted to segment other organs.Source: CONCLUSION · p. 14
• The network is reported to handle irregular prostates and rectums and achieve higher DSC than state-of-the-art methods, including the authors' 2D model.Source: CONCLUSION · p. 14
• The authors state that the workflow could be adapted for segmentation of other organs.Source: CONCLUSION · p. 14

Appendix A. Details on the 5 channel 2D U-Net architectures used for organ localization

• A passage labels an appendix section as details on modified 3D U-Net architectures for organ segmentation.Source: Appendix A · p. 18
• The listed architecture fragments include four Max Pooling labels.Source: Appendix A · p. 19
• Other fragments list Conv labels 8 and 1, and name a learning curve for a single prostate validation set.Source: Appendix A · p. 20