PADLLS

PADLLS performs automated liver segmentation from contrast-enhanced CT scans using cascaded deep convolutional neural networks to reduce oversegmentation and produce high-precision liver masks for image-based analysis.

Key Features:

• Automated workflow: Employs a cascading DCNN approach where a V-net model generates initial rough masks for the liver, spleen, and stomach.
• Refinement process: Excludes non-liver regions using spleen and stomach masks and removes ascites via a morphological algorithm to refine the scan.
• Final segmentation: Processes the refined image with an H-DenseUnet model to produce the final liver segmentation, minimizing oversegmentation.

Scientific Applications:

• Clinical image analysis: Enables accurate liver segmentation for diagnosis, treatment planning, and monitoring of liver diseases from contrast-enhanced CT scans.

Methodology:

Trained on over 200 CT scans and evaluated against SLIVER07 and 3DIRCADb benchmark datasets; reported Dice scores: SLIVER07 0.957 ± 0.033 (vs H-DenseUnet 0.927 ± 0.044 and V-net 0.872 ± 0.121) and 3DIRCADb 0.965 ± 0.016 (vs H-DenseUnet 0.930 ± 0.041 and V-net 0.874 ± 0.060); pipeline steps explicitly include V-net initial segmentation, exclusion of non-liver masks, ascites removal via a morphological algorithm, and final H-DenseUnet segmentation.