Deep feature fusion and optimized feature selection based ensemble classification of liver lesions

Classification of liver abnormalities is crucial for the early identification of liver cancer. In clinical settings, radiological professionals typically make diagnoses manually which is subjective, time-consuming and vulnerable to error. Therefore, there is still a demand for precise classification of liver diseases. We propose an Ensemble learning-based classification model to classify liver lesions on CT images. To excerpt all the essential facts from the image, deep feature fusion is incorporated by concatenating the features from pre-trained deep CNN models densenet201 and InceptionResnetV2. To minimize the feature space and boost classification accuracy, hybrid optimization methodologies, Genetic Algorithms and Ant Colony Optimization are applied. Finally, a Heterogeneous Ensemble classifier divides the retrieved features into four groups (liver abscess, liver cirrhosis, hepatocellular carcinoma, and metastasis). It is clearly seen and observed that 98.3% accuracy is contributed by ensemble classifier with the support of concatenated deep features and this classifier excels in all other ways and means.

Automated summary

The classification of liver abnormalities is essential for early detection of liver cancer. Manual diagnoses by radiological professionals in clinical settings are subjective, time-consuming, and prone to errors. Therefore, there is a need for precise classification of liver diseases. We propose an ensemble learning-based classification model that incorporates deep feature fusion and hybrid optimization methodologies to classify liver lesions on CT images. The heterogeneous ensemble classifier achieves an accuracy of 98.3% by utilizing concatenated deep features and outperforms other methods.