Medical machine intelligence: swarm optimization, feature fusion, and neighboring-awareness

Medical imaging can be one of the most significant bridges that connect clinicians to patients because medical imaging can produce clear visualization results on human organs under the skin so that clinicians can get a clear understanding of the structure of the organs and conduct accurate analysis. The medical images contain massive information, but manual interpretation often suffers from high inter-observer variance and low reproducibility. In addition, image analysis poses a heavy burden on physicians and surgeons.

In this thesis, my research on automatic medical image interpretation is presented based on deep learning and computer vision. Specifically, the contributions to the detection of COVID-19 using chest CT scans are demonstrated. Firstly, swarm optimization is employed to find the best parameters in the proposed deep model for COVID-19 detection. An optimal backbone selection method is proposed to obtain the best backbone model, and an extreme learning machine is used for classification. A bat algorithm with chaotic maps is presented to further optimize the weights and biases in the hidden layer of the extreme learning machine. Experimental results reveal the effectiveness of the proposed model CoDeNet.

However, swarm optimization is time-consuming in applications because classifiers can be trained more frequently than backbone models in real-world applications. Therefore, a feature fusion mechanism is proposed for automatic chest CT analysis. Two different pre-trained CNN models are fine-tuned on the CT dataset for image feature extraction. The two feature sets are then fused together by discriminant correlation analysis. Three RNNs are trained as the classifier in the proposed model, and the predictions of the three RNNs are fused by majority voting to produce the final predictions of the proposed model DFFNet-M.

The fusion of multiple features and classifiers is effective for COVID-19 diagnosis, but the context information within a single feature set can be further explored. The classification performance of the CAD system can be improved by integrating the neighboring information among the samples to improve. The graph theory is introduced to embed the neighboring information among the feature vectors, and a neighboring-aware graph neural network is presented for COVID-19 classification. A graph random vector functional-link net is trained as the classifier in the model. Experiment results from 5-fold cross-validation show that the proposed NAGNN can achieve better classification performance than the baseline models.