Flexible ConvNext Block Based Multi-task Learning Framework for Liver MRI Images Analysis

Dong Sui, Weifeng Liu, Maozu Guo, Yang Li, Lei Zhang, Gongning Luo, Kuanquan Wang

Abstract: Liver cancer is the second leading cause of death all over the world in the 2020s’, and the incidence rate has been growing on a global scale and become a serious threat to human life. Early diagnosis of liver cancer from medical images can allow the patients to receive better treatment and achieve good outcomes. Although medical imaging approaches have made significant progress over the past decades, there are still great demands to reconstruct the network structure for the adaptation to downstream tasks. It remains a great challenge for liver tumor identification from MRI images. Recently, self-attention mechanism based transformer models can capture long-range dependencies, which make them perform well on many medical image analysis tasks. Such as Segformer and TransUNet, since lacking the translation in-variance and inductive bias of CNNs, they are still needed large-scale training to fill the gap, especially in the field of medical image analysis domain. In this study, we incorporate a novel flexible Condeathblock as a feature extractor to perform liver images analysis and propose a new analytical framework CXNet to extract discriminative multi-scale visual representations. The experiments results demonstrated our framework outperforms the state-of the-art models on three datasets, including a publicly available liver segmentation dataset as well as two in-house liver tumor classification/segmentation datasets. On the 3DIRCADb dataset, our CXNet outperforms the UNet model by 2.82%, 2.73%, and 4.46% in terms of the Jaccard similarity coefficient, Dice coefficient, and accuracy, and outperforms the TransUNet model by 8.49%, 5.35%,which 3.62%, respectively. Code is available at https://github.com/SPECTRELWF/CXNet