A semantic segmentation framework for liver and liver tumour segmentation

Semantic segmentation is a critical computer vision task with numerous real-world applications. While U-Net+ + excels at capturing multi-scale contextual information, it has limitations, like any deep learning model. This research integrates DenseNet, Atrous spatial pyramid pooling (ASPP), and U-Net+ + to address these limitations. This enhancement is particularly beneficial for precise and detailed segmentation tasks. The inclusion of DenseNet enables U-Net+ + to capture even the most subtle data details, which is essential for complex or highly detailed segmentation tasks requiring accurate object boundary delineation. To assess the practical effectiveness of our method, we conducted a real-world evaluation using a novel dataset, ATC-BATNA2020, collected from the Anti-Tumour Centre in Batna, Algeria. This evaluation not only confirms the theoretical robustness of our approach, but also demonstrates its effectiveness in clinical and real-world scenarios. Our experiments on the ATC-BATNA2020 dataset and the well known existing LiTS17 dataset yielded high accuracies: 99.28% and 99.23% for liver segmentation, and 97.58% and 97.41% for liver tumour segmentation respectively. These results indicate that our proposed method achieves superior accuracy and outperforms other leading methodologies.