Gul Sidra, Khan Muhammad Salman, Hossain Md Sakib Abrar, Chowdhury Muhammad E H, Sumon Md Shaheenur Islam
Department of Computer Systems Engineering, University of Engineering and Technology, Peshawar 25000, Pakistan.
Artificial Intelligence in Healthcare, Intelligent Information Processing Lab, National Center of Artificial Intelligence, Peshawar 25000, Pakistan.
Diagnostics (Basel). 2024 Dec 8;14(23):2761. doi: 10.3390/diagnostics14232761.
Accurate liver and tumor detection and segmentation are crucial in diagnosis of early-stage liver malignancies. As opposed to manual interpretation, which is a difficult and time-consuming process, accurate tumor detection using a computer-aided diagnosis system can save both time and human efforts. We propose a cascaded encoder-decoder technique based on self-organized neural networks, which is a recent variant of operational neural networks (ONNs), for accurate segmentation and identification of liver tumors. The first encoder-decoder CNN segments the liver. For generating the liver region of interest, the segmented liver mask is placed over the input computed tomography (CT) image and then fed to the second Self-ONN model for tumor segmentation. For further investigation the other three distinct encoder-decoder architectures U-Net, feature pyramid networks (FPNs), and U-Net++, have also been investigated by altering the backbone at the encoders utilizing ResNet and DenseNet variants for transfer learning. For the liver segmentation task, Self-ONN with a ResNet18 backbone has achieved a dice similarity coefficient score of 98.182% and an intersection over union of 97.436%. Tumor segmentation with Self-ONN with the DenseNet201 encoder resulted in an outstanding DSC of 92.836% and IoU of 91.748%. The suggested method is capable of precisely locating liver tumors of various sizes and shapes, including tiny infection patches that were said to be challenging to find in earlier research.
准确的肝脏和肿瘤检测与分割对于早期肝脏恶性肿瘤的诊断至关重要。与手动解读这一困难且耗时的过程不同，使用计算机辅助诊断系统进行准确的肿瘤检测可以节省时间和人力。我们提出一种基于自组织神经网络的级联编码器 - 解码器技术，自组织神经网络是运算神经网络（ONNs）的一种最新变体，用于准确分割和识别肝脏肿瘤。第一个编码器 - 解码器卷积神经网络（CNN）对肝脏进行分割。为了生成感兴趣的肝脏区域，将分割后的肝脏掩码放置在输入的计算机断层扫描（CT）图像上，然后将其输入到第二个自组织运算神经网络（Self - ONN）模型进行肿瘤分割。为了进一步研究，还通过在编码器处改变骨干网络，利用ResNet和DenseNet变体进行迁移学习，研究了其他三种不同的编码器 - 解码器架构，即U - Net、特征金字塔网络（FPN）和U - Net++。对于肝脏分割任务，具有ResNet18骨干网络的Self - ONN的骰子相似系数得分达到了98.182%，交并比为97.436%。使用具有DenseNet201编码器的Self - ONN进行肿瘤分割，得到了出色的92.836%的DSC和91.748%的IoU。所提出的方法能够精确地定位各种大小和形状的肝脏肿瘤，包括在早期研究中被认为难以发现的微小感染斑块。
