A Multi-Scale Residual Learning-Based Sequential Model for Hepatocellular Carcinoma Detection
TL;DR
An efficient HCC diagnosis model called MS-ResNet is designed by combining the proficiencies of deep learning and CT imaging in its process to detect HCC, but unlike other approaches, the new method can take advantage of multi-scale features to improve the overall efficiency of the diagnostic framework with the DL technique.
Abstract
Accurate recognition of Hepatocellular Carcinoma (HCC) from medical images is very important to diagnose the disease early and come up with appropriate treatment plans. For instance, in this paper, an efficient HCC diagnosis model called MS-ResNet is designed by combining the proficiencies of deep learning. In particular, the proposed approach uses CT imaging in its process to detect HCC, but unlike other approaches, the new method can take advantage of multi-scale features to improve the overall efficiency of the diagnostic framework with the DL technique. It involves an improved preprocessing step where normalization and CLAHE techniques are used to enhance the quality of the input data. Further, in this work, multi-scale features have been extracted via an enhanced residual network. These extracted features have been further modeled through sequences using a Bi-LSTM approach and attention mechanisms. Subsequently, the predictive capability of the model in recognizing HCC is tested through the following evaluation metrics: accuracy, precision, recall, F1-score, and Dice score. From the experimental analysis of results presented in the work, the proposed method achieves an accuracy of 96.40%, precision of 95.70%, recall of 96.40%, F1-score of 94.90%, and dice score of 94.90%. These results exceed those obtained by other classifiers such as CNN, VGGNet, and ResNet, and are compared with the several recently developed liver tumor segmentation and classification methods.