CUNet-CLSTM: A Novel Fusion of CUNet and CLSTM for Superior Liver Cancer Detection in CT Scans

Abstract

Liver cancer remains one of the main causes of mortality worldwide, requiring advancements in diagnostic systems for early detection and improved patient outcomes. Computed tomography (CT) scans are crucial in the diagnosis of liver cancer, providing detailed images for medical analysis. Despite the triumph of convolutional neural networks in medical image analysis, challenges such as overfitting, limited labeled data, and complex tumor morphology hinder accurate detection of liver cancer. This study proposes a novel architecture, the cascaded UNet convolutional long short-term model (CUNet-CLSTM), which leverages the strengths of UNet and convolutional long short-term memory to improve liver segmentation and tumor detection in CT images. The anticipated CUNet-CLSTM model evaluated on two benchmark datasets: LiTS and 3DICARDB. The experimental results show excellent performance, achieving the best DICE values of 91.3% for liver segmentation and 96.1% for tumor detection on the LiTS dataset, and 96.3% for liver segmentation and 98.3% for tumor detection on the 3DICARDB dataset. Compared to state-of-the-art models, the proposed CUNet-CLSTM not only improves segmentation accuracy, but also mitigates overfitting through sequential feature learning and optimizes computational efficiency by providing an end-to-end framework, reducing the complexity of multi-stage pipelines. This research contributes to addressing the urgent need for accurate methods for detecting liver cancer, facilitating early diagnosis, and customized treatment strategies.