Enterprise AI Analysis: Deep Learning for Automatic Segmentation of Hepatocellular Carcinoma

The study compares three prominent 3D deep learning architectures: nnU-Net (a self-configuring convolutional network with Residual Encoder and Dice + Binary Cross-Entropy loss), SwinUNETR (an encoder-decoder with Swin Transformer and CNNs, using Dice Cross-Entropy loss and pre-training), and U-Mamba (a hybrid CNN-SSM network, trained with Dice + Binary Cross-Entropy loss in a bottleneck configuration). These models were trained and tested on four datasets: LiTS, HCC-TACE-Seg, WAW-TACE, and the newly introduced HCC-ARSeg dataset. The evaluation primarily focused on the portal-venous phase and used Dice Score and Symmetric Average Surface Distance (SASD) as metrics.

The evaluation utilized four datasets:
LiTS: A public benchmark with diverse liver lesions from seven sites.
HCC-TACE-Seg: 105 CT scans of HCC patients before TACE, standardized acquisition.
WAW-TACE: 233 multi-phase CT images of treatment-naive HCC patients.
HCC-ARSeg: A novel proprietary dataset of 94 HCC patients with contrast-enhanced CT scans across arterial, portal-venous, and delayed phases, meticulously annotated by experts. This dataset uniquely provides all three contrast phases with validated segmentations for HCC, addressing a critical gap in existing resources.

The nnU-Net architecture consistently outperformed U-Mamba and SwinUNETR across all evaluated datasets in terms of Dice score and SASD. While U-Mamba showed comparable Dice, nnU-Net achieved superior boundary precision. SwinUNETR performed significantly lower, likely due to limited dataset size for its transformer-based design. The WAW-TACE dataset yielded the best overall results for nnU-Net (Dice: 0.782, SASD: 2.720). The novel HCC-ARSeg dataset, despite having the lowest tumor pixel percentage (0.06%), demonstrated strong performance with nnU-Net (Dice: 0.752, SASD: 4.680), and its modified ROI-based version achieved excellent SASD (3.903). The multi-phasic nature and expert annotation of HCC-ARSeg are crucial for advancing HCC classification.