Adaptive Feature Engineering Modeling for Ultrasound Image Classification for Decision Support
Date of Award
Doctor of Philosophy (PhD)
Computational Analysis and Modeling
Ultrasonography is considered a relatively safe option for the diagnosis of benign and malignant cancer lesions due to the low-energy sound waves used. However, the visual interpretation of the ultrasound images is time-consuming and usually has high false alerts due to speckle noise. Improved methods of collection image-based data have been proposed to reduce noise in the images; however, this has proved not to solve the problem due to the complex nature of images and the exponential growth of biomedical datasets. Secondly, the target class in real-world biomedical datasets, that is the focus of interest of a biopsy, is usually significantly underrepresented compared to the non-target class. This makes it difficult to train standard classification models like Support Vector Machine (SVM), Decision Trees, and Nearest Neighbor techniques on biomedical datasets because they assume an equal class distribution or an equal misclassification cost. Resampling techniques by either oversampling the minority class or under-sampling the majority class have been proposed to mitigate the class imbalance problem but with minimal success. We propose a method of resolving the class imbalance problem with the design of a novel data-adaptive feature engineering model for extracting, selecting, and transforming textural features into a feature space that is inherently relevant to the application domain.
We hypothesize that by maximizing the variance and preserving as much variability in well-engineered features prior to applying a classifier model will boost the differentiation of the thyroid nodules (benign or malignant) through effective model building. Our proposed a hybrid approach of applying Regression and Rule-Based techniques to build our Feature Engineering and a Bayesian Classifier respectively.
In the Feature Engineering model, we transformed images pixel intensity values into a high dimensional structured dataset and fitting a regression analysis model to estimate relevant kernel parameters to be applied to the proposed filter method. We adopted an Elastic Net Regularization path to control the maximum log-likelihood estimation of the Regression model. Finally, we applied a Bayesian network inference to estimate a subset for the textural features with a significant conditional dependency in the classification of the thyroid lesion. This is performed to establish the conditional influence on the textural feature to the random factors generated through our feature engineering model and to evaluate the success criterion of our approach.
The proposed approach was tested and evaluated on a public dataset obtained from thyroid cancer ultrasound diagnostic data. The analyses of the results showed that the classification performance had a significant improvement overall for accuracy and area under the curve when then proposed feature engineering model was applied to the data. We show that a high performance of 96.00% accuracy with a sensitivity and specificity of 99.64%) and 90.23% respectively was achieved for a filter size of 13 × 13.
Mugasa, Hatwib, "" (2019). Dissertation. 822.
Applied Mathematics Commons, Computer Sciences Commons, Statistics and Probability Commons