Publication

Abstract : Accurate and timely diagnosis of liver lesions is essential for enhancing clinical decisions. Ultrasound imaging is often limited by low contrast and speckle noise significantly impacting the quality of extracted diagnostic features. However, the quality of diagnosis mainly depends on the skill of the sonographer in interpreting liver lesions. Currently, deep learning based diagnosis systems depend on the quality and consistency of medical imaging data. Therefore, to mitigate these challenges, this article presents a computer-aided diagnostic framework for diagnosing focal liver lesion (FLL) images. The proposed framework utilizes DarkNet-19 as a backbone model to extract activation attributes from its first convolutional block. These activation cues are further enhanced by a bottleneck attention module that adaptively refines the salient channel and spatial information to improve the quality of feature representation. Further, the refined feature maps are processed through the remaining convolutional layers, followed by the global average pooling layer, a fully connected layer with three output nodes, and a softmax classification layer for the final prediction of FLL classes. Experimental analysis of the SYSU-FLL-CEUS dataset demonstrates that the proposed framework surpasses conventional and recent approaches for FLL classification across multiple performance metrics, showcasing its potential for advancing deep learning applications in liver disease diagnostics.