INNOVATIVE FEATURE EXTRACTION AND ATTENTION MECHANISMS IN COMPUTER AIDED DIAGNOSIS (CAD) MODELS FOR DISEASE PREDICTION

Abstract

In medical imaging informatics, the development of computer-aided detection and/or diagnosis (CAD) schemes has been an active research issue with promising outcomes in helping doctors make better clinical and/or diagnostic decisions in the last 20 years. To create strong CAD schemes, we must create cutting-edge machine learning (ML) and image processing algorithms that maximize every stage of the CAD pipeline, from feature generation to detection and segmentation of the region of interest to integration with ML classifiers. In this report, I conducted multiple studies investigating the feasibility of fine-tuning machine learning and deep learning models for accurate prediction of diseases. The first study aims to improve the performance of computer-aided diagnosis scheme to improve the performance of breast lesion classification scheme using a new feature fusion method. A dataset with 2,000 mammograms was assembled. Based on lesion center, a Region of Interest (ROI) extracted from each image. Among them, 1,000 ROIs depict verified malignant lesions and the rest include benign lesions. An adaptive multilayer region growing algorithm is applied to segment suspicious lesions. Several sets of statistical features, texture features based on Gray-Level Run-Length Matrix (GLRLM), Gray Level Dependence Matrix (GLDM), Gray Level Cooccurrence Matrix (GLCM), Wavelet transformed features and shape-based features are computed from the original ROI and segmented lesion, respectively. Three support vector machines (SVM) are trained using features computed from original ROIs, segmented lesions, and fusion of both, respectively, using a 10-fold cross-validation. By applying the area under the ROC curve (AUC) as an evaluation index, my study results reveal no significant difference between AUC values computed using classification scores generated by two SVMs trained with features computed from original ROIs or segmented lesions. However, utilizing the fused features, AUC of SVM increases more than 10%. The study also demonstrates that the unpaired t-student tests could be used to select effective features from the fusion process. The objective of the second study is to investigate the feasibility of identifying and applying a new feature extraction technique to predict the survival of gastric cancer patients. A retrospective dataset including the computed tomography (CT) images of 135 patients was assembled. Among them, 68 patients survived longer than three years. Several sets of radiomics features were extracted and were incorporated into a machine learning model, and their classification performance was characterized. To improve the classification performance, another 27 texture and roughness parameters with 2484 superficial and spatial features were extracted to propose a new feature pool. This new feature set was added into the machine learning model and its performance was analyzed. To determine the best model for the experiment, Random Forest (RF) classifier, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naïve Bayes (NB) (four of the most popular machine learning models) were utilized. The models were trained and tested using the five-fold cross-validation method. Using the area under ROC curve (AUC) as an evaluation index, the model that was generated using the new feature pool yields AUC = 0.98 ± 0.01, which was significantly higher than the models created using the traditional radiomics feature set (p < 0.04). RF classifier performed better than the other machine learning models. This study demonstrated that although radiomics features produced good classification performance, creating new feature sets significantly improved the model performance. The purpose of the third study is to optimize a new deep transfer learning model by implementing a novel attention mechanism to improve the accuracy of breast lesion classification. ResNet50 was selected as the base model to develop a new deep transfer learning model. To enhance the accuracy of breast lesion classification, a convolutional block attention module (CBAM) was added to the standard ResNet50 model and optimizing a new model for this task. A large dataset with 4280 mammograms depicting suspicious soft-tissue mass-type lesions was assembled. A ROI is extracted from each image based on lesion center. Among them, 2480 and 1800 ROIs depict verified benign and malignant lesions, respectively. The image dataset is randomly split into two subsets with a ratio of 9:1 five times to train and test two ResNet50 models with and without using CBAM. Using AUC as an evaluation index, the new CBAM-based ResNet50 model yields AUC = 0.866 ± 0.015, which is significantly higher than that obtained by the standard ResNet50 model (AUC = 0.772 ± 0.008) (p < 0.01). This study demonstrates that although deep transfer learning technology attracted broad research interest in medical-imaging informatic fields, adding a new attention mechanism to optimize deep transfer learning models for specific application tasks can play an important role in further improving model performances. In summary, in all these studies, I developed and investigated several key concepts of the improving of machine learning and deep learning techniques including (i) feature extraction, and (ii) novel attention mechanism. All developed CAD models are embedded with interactive and visually aided graphical user interfaces (GUIs) to provide user functionality. These techniques present innovative approaches for building efficient ML and DL models. The study results indicate the underlying CAD scheme's potential application to assist radiologists in clinical settings for their assessments in diagnosing disease and improving their overall performance.