Performance analysis of artificial intelligence–driven convolutional neural network architectures for liver tumor segmentation.

Abstract

e16307 Background: Advances in medical image segmentation have raised debate about the practical performance of the latest architectures and CNN-based approaches. Recent studies have demonstrated that CNN-based architectures maintain competitive performance compared to newer architectural paradigms despite limited task-specific validation. Methods: This study utilizes the ATLAS dataset from the MICCAI 2023 ATLAS Challenge, including a total of 90 pairs of MRI images and labels with varieties of machines, sequences, and contrast phases. This dataset was divided into a public set of 60 patients and a private set of 30 patients collected between 2012-2023. Due to the variety of images, stratified k-fold cross-validation is applied. STU-Net family is leveraged for evaluation and comparison with U-Net variants as baselines. All models are trained using the nnU-Net framework with default combined loss Dice and Cross Entropy. For a fair comparison, all metrics from the ATLAS Challenge are calculated twice to evaluate our models, once for a whole liver and once for tumor segmentation. Results: Our results on the ATLAS dataset revealed that while traditional U-Net variants establish strong baseline performance, STU-Net achieves superior capabilities across various evaluation metrics, notably dice scores of 95.76 ± 0.99% and 68.30 ± 2.13% for liver and tumor segmentation, respectively. These results validate its effectiveness for complex medical segmentation tasks. Conclusions: These results affirm the effectiveness of CNN-based models, particularly STUNet, in complex hepatic tumor segmentation tasks. STUNet’s advanced feature extraction and scalability make it a promising solution for automated liver tumor segmentation in clinical practice. Future work will focus on enhancing interpretability and clinical integration.