Objective: Radiomics and deep transfer learning are two popular technologies used to develop computer-aided detection and diagnosis (CAD) schemes of medical images. This study aims to investigate and to compare the advantages and the potential limitations of applying these two technologies in developing CAD schemes. Methods: A relatively large and diverse retrospective dataset including 3000 digital mammograms was assembled in which 1496 images depicted malignant lesions and 1504 images depicted benign lesions. Two CAD schemes were developed to classify breast lesions. The first scheme was developed using four steps namely, applying an adaptive multi-layer topographic region growing algorithm to segment lesions, computing initial radiomics features, applying a principal component algorithm to generate an optimal feature vector, and building a support vector machine classifier. The second CAD scheme was built based on a pre-trained residual net architecture (ResNet50) as a transfer learning model to classify breast lesions. Both CAD schemes were trained and tested using a 10-fold cross-validation method. Several score fusion methods were also investigated to classify breast lesions. CAD performances were evaluated and compared by the areas under the ROC curve (AUC). Results: The ResNet50 model-based CAD scheme yielded AUC = 0.85 ± 0.02, which was significantly higher than the radiomics feature-based CAD scheme with AUC = 0.77 ± 0.02 (p < 0.01). Additionally, the fusion of classification scores generated by the two CAD schemes did not further improve classification performance. Conclusion: This study demonstrates that using deep transfer learning is more efficient to develop CAD schemes and it enables a higher lesion classification performance than CAD schemes developed using radiomics-based technology.

Objective: Radiomics and deep transfer learning are two popular technologies used to develop computer-aided detection and diadnosis (CAD) schemes of medical images. This study aims to investigate and compare advantages and potential limitations of applying these two technologies in developing CAD schemes. Methods: A relatively large and diverse retrospective dataset including 3,000 digital mammograms is assembled in which 1,496 images depict malignant lesions and 1,054 images depict benign lesions. Two CAD schemes are developed to classify breast lesions. The first scheme is developed using four steps namely, applying an adaptive multi-layer topographic region growing algorithm to segment lesions, computing initial radiomics features, applying a principal component algorithm to generate an optimal feature vector, and building a support vector machine classifier. The second CAD scheme is built based a pre-trained residual net architecture (ResNet50) as a transfer learning model to classify breast lesions. Both CAD schemes are trained and tested using a 10-fold cross-validation method. Several score fusion methods are also investigated to classify breast lesions. CAD classification performance is evaluated by the area under ROC curve (AUC). Results: ResNet50 model-based CAD scheme yields AUC = 0.85±0.02, which is significantly higher than radiomics feature-based CAD scheme with AUC = 0.77±0.02 (p < 0.01). Additionally, fusion of classification scores generated by two CAD schemes does not further improve classification performance. Conclusion: This study indicates that using deep transfer learning is more efficient to develop CAD schemes and enables to yield higher lesion classification performance than CAD schemes developed using radiomics-based technology.

Deep Convolutional Neural Networks (DCNNs) have emerged as powerful components of computer-aided systems to detect and classify diseases in medical images. In order to robustly achieve high performance, a large and diverse dataset with annotated images is required to train or fine-tune DCNNs. However, such datasets are often very difficult and/or expensive to obtain due to demands on clinicians’ time and inter-reader variability. To address these challenges and increase training dataset size, we propose an efficient multi-stage algorithm to generate synthetic medical image data by extracting annotated diseased regions and randomly projecting them onto disease-free images. To test the feasibility of this new algorithm, the publicly available Indian Diabetic Retinopathy Image Dataset (IDRiD) is used. This dataset is comprised of the annotated fundus images acquired from 81 patients with two categories of diseases. Among them, 54 and 27 images are used for training and testing images, respectively. Using the proposed algorithm, we generate synthetic data by inserting extracted diseased lesions onto another set of 60 disease-free images, which results in 7,902 and 6,786 images for the two categories of diseases, respectively. Three transfer-learning-based DCNN models (VGG16, ResNet50, and Inception-v3) are trained using original IDRiD images and synthetic dataset, respectively. When applied to the same test images, the model trained with the synthetic dataset outperformed the model trained using he original IDRiD dataset by 7.4% in disease classification. These results indicate that this simple algorithm can generate diverse and useful synthetic lesion data to improve performance of DCNN models.

Machine learning is widely used in developing computer-aided diagnosis (CAD) schemes of medical images. However, CAD usually computes large number of image features from the targeted regions, which creates a challenge of how to identify a small and optimal feature vector to build robust machine learning models. In this study, we investigate feasibility of applying a random projection algorithm to build an optimal feature vector from the initially CAD-generated large feature pool and improve performance of machine learning model. We assemble a retrospective dataset involving 1,487 cases of mammograms in which 644 cases have confirmed malignant mass lesions and 843 have benign lesions. A CAD scheme is first applied to segment mass regions and initially compute 181 features. Then, support vector machine (SVM) models embedded with several feature dimensionality reduction methods are built to predict likelihood of lesions being malignant. All SVM models are trained and tested using a leave-one-case-out cross-validation method. SVM generates a likelihood score of each segmented mass region depicting on one-view mammogram. By fusion of two scores of the same mass depicting on two-view mammograms, a case-based likelihood score is also evaluated. Comparing with the principle component analyses, nonnegative matrix factorization, and Chi-squared methods, SVM embedded with the random projection algorithm yielded a significantly higher case-based lesion classification performance with the area under ROC curve of 0.84+0.01 (p<0.02). The study demonstrates that the random project algorithm is a promising method to generate optimal feature vectors to help improve performance of machine learning models of medical images.