Automated Liver Lesion Segmentation with Deep Learning Model

Proper segmentation of liver lesions is important with regards to computer aided diagnosis and treatment by means of hepatic diseases. Nevertheless, the current models can fail in identification of tiny, poor-contrast, and asymmetric lesions that might have limited feature representation or over-expensive computation. In this paper, a new and efficient deep convolutional neural network (DCNN) architecture is proposed that uses depth wise separable convolutions so that the feature extraction ability can be improved, meanwhile, the model is highly simplified. The model includes batch normalization and dropout layers that have been integrated to enhance stability and generalizability in the data-scarce medical settings regarding training. It is a sparse architecture with effective performance in capturing the finer-grained details of a lesion. Testing on the public LiTS dataset proves the better performance of the model which attains a Dice Coefficient of 0.8209, IoU of 0.7007, Precision of 0.8454, Recall of 0.8054, and F1 Score of 0.8209 and the test loss is low with a value of 0.2095. The overall performance is highly beneficial with better segmentation accuracy and the accuracy of delineating the boundaries compared with conventional models; hence comparative analysis. The framework is suitable and can be implemented in real-time in clinical decision-support systems because it has a potential trade-off between the performance and the efficiency. This is a contribution in terms of a solid and scale computation in liver lesion segmentation, which is a contribution to the application of deep learning to AI in radiology.