Locating tumors from medical images is of high importance in medical analysis and diagnosis. To tackle the complicated shape of tumors, we propose a multi-leves.l feature extraction neural network to automatically segment the data. Our proposed model is trained and tested with one liver tumor ultrasound and two CT datasets. We employ ++, a collaborative model that uses modified nested U-Net, as our backbone. The model is integrated with dilated dense short skip connections within convolution blocks to further improve the gradient flow and feature preservation. In addition, we modify the original Atrous Spatial Pyramid Pooling (ASPP) to an adaptive pooling structure for better compatibility with nested U-Net. Adaptive ASPP is designed to extract features from different levels and cover the increasing range of feature extraction with regard to the depth of the nested network. Our model showed its advantage in accurately segmenting different tumor sizes with complex edges and was able to generalize with small and diverse datasets. We further improved our model with the newly introduced AdaBelief optimizer and achieved a faster convergence rate. Segmentation results showed that the proposed model outperformed multiple network structures, and achieved a 0.9153 dice coefficient for the ultrasound dataset, a 0.9413 and a 0.9246 dice coefficient for the two CT datasets.
All Science Journal Classification (ASJC) codes
- Computer Science Applications
- Health Informatics