This study aims to enable effective breast ultrasound image classification by combining deep features with conventional handcrafted features to classify the tumors. In particular, the deep features are extracted from a pre-trained convolutional neural network model, namely the VGG19 model, at six different extraction levels. The deep features extracted at each level are analyzed using a features selection algorithm to identify the deep feature combination that achieves the highest classification performance. Furthermore, the extracted deep features are combined with handcrafted texture and morphological features and processed using features selection to investigate the possibility of improving the classification performance. The cross-validation analysis, which is performed using 380 breast ultrasound images, shows that the best combination of deep features is obtained using a feature set, denoted by $CONV$ features that include convolution features extracted from all convolution blocks of the VGG19 model. In particular, the $CONV$ features achieved mean accuracy, sensitivity, and specificity values of $94.2\%$, $93.3\%$, and $94.9\%$, respectively. The analysis also shows that the performance of the $CONV$ features degrades substantially when the features selection algorithm is not applied. The classification performance of the $CONV$ features is improved by combining these features with handcrafted morphological features to achieve mean accuracy, sensitivity, and specificity values of $96.1\%$, $95.7\%$, and $96.3\%$, respectively. Furthermore, the cross-validation analysis demonstrates that the $CONV$ features and the combined $CONV$ and morphological features outperform the handcrafted texture and morphological features as well as the fine-tuned VGG19 model. The generalization performance of the $CONV$ features and the combined $CONV$ and morphological features is demonstrated by performing the training using the 380 breast ultrasound images and the testing using another dataset that includes 163 images. The results suggest that the combined $CONV$ and morphological features can achieve effective breast ultrasound image classifications that increase the capability of detecting malignant tumors and reduce the potential of misclassifying benign tumors.

Breast ultrasound (BUS) imaging is commonly used for breast cancer diagnosis, but the interpretation of BUS images varies based on the radiologist's experience. Computer-aided diagnosis (CAD) systems have been proposed to provide the radiologist with an objective, computer-based classification of BUS images. Nevertheless, the majority of these systems are based on handcrafted features that are designed manually to quantify the tumor. Hence, the accuracy of these CAD systems depends on the capability of the handcrafted features to differentiate between benign and malignant tumors. Convolutional neural networks (CNNs) provide a promising approach to improve the classification of BUS images due to their ability to achieve data-driven extraction of objective, accurate, and generalizable image representations. However, the limited size of the available BUS image databases might restrict the capability of training the CNNs from scratch. To address this limitation, we investigate the use of two approaches, namely the deep features extraction approach and transfer learning approach, to enable the use of a pre-trained CNN model to achieve accurate classification of BUS images. The results show that the deep features extraction approach outperforms the transfer learning approach. Moreover, the results indicate that the extraction of deep features from the pre-trained CNN model, which is combined with effective features selection, has enabled accurate BUS image classification with accuracy, sensitivity, and specificity values of 93.9%, 95.3%, and 92.5%, respectively. These results suggest the feasibility of combining deep features extracted from pre-trained CNN models with effective features selection algorithms to achieve accurate BUS image classification.

In electrocardiography (ECG), measurement of QRS duration (QRSd) is crucial for diagnosing conditions such as left bundle branch block. To address the limited availability of ECG databases with QRS delineation labels, we present a method to use small databases to train deep learning object detection models for global QRSd estimation that involves minimal manual labeling of median beats. In our method, an ECG record is segmented into individual heartbeats, transformed into artificial images, and a Faster R-CNN model is utilized to estimate the global QRSd. Faster R-CNN models were tested with three different backbone configurations (VGG-16, VGG-19, and RESNET-18) and two ECG image formats: binary images in which each beat in each lead was represented by a separate image and RGB images in which the same beat from a trio of leads was superimposed by mapping each lead to a different color channel. Using 258 twelve-lead, 10-s digital ECG records acquired from 140 unique heart failure outpatients, the best-performing backbone, VGG-19 with RGB images, achieved root-mean-square and mean absolute errors for QRSd of 10.4 ± 0.8 ms and 8.2 ± 1.0 ms, respectively, during five-fold cross-validation. Testing with an independent, publicly available dataset yielded root-mean-square and mean absolute errors for QRSd of 7.0 ± 1.1 ms and 5.3 ± 0.9 ms, respectively. Therefore, our method provides high QRSd estimation accuracy while reducing the need for manual labeling and shows promise for generalization to independent databases, demonstrating potential for efficient training of deep learning models on small ECG databases.

Novel coronavirus disease 2019 (COVID-19) is highly contagious and can lead to serious medical complications. Early detection of COVID-19 is important to control the spread of the disease and reduce the associated mortality rates. The reverse transcription polymerase chain reaction (RT-PCR) is commonly used for COVID-19 diagnosis. However, the RT-PCR is time consuming, requires special materials, and might have limited detection sensitivity in mild cases. One of the promising complementary modalities to improve the detection and tracking of COVID-19 is X-ray imaging of the chest, but the task of interpreting chest X-ray images is challenging. Convolutional neural networks (CNNs) provide an effective computational tool for classifying chest X-ray images with the goal of achieving accurate COVID-19 diagnosis. This study investigates the application of two pre-trained CNN models, namely AlexNet and ResNet-50, using transfer learning to classify chest X-ray images as normal, pneumonia (non-COVID-19 pneumonia), and COVID-19. The transfer learning process was applied based on two classification approaches, which are the flat classification approach and the hierarchical classification approach. The performance of the proposed CNN-based classification schemes has been evaluated using a dataset that includes 8,703 chest X-ray images. The results indicate that the pre-trained CNN models combined with the hierarchical classification approach achieved effective classification of chest X-ray images. In particular, the pre-trained AlexNet model that is combined with the hierarchical classification approach obtained macro-averaged classification specificity, sensitivity, and F1 score of 98.3%, 89.1%, and 91.4%, respectively. Furthermore, the pre-trained ResNet-50 model that is combined with the hierarchical classification approach achieved macro-averaged specificity, sensitivity, and F1 score of 97.4%, 95.2%, and 94.9%, respectively.