Mgs M Luthfi Ramadhan

Heart failure is a disease many consider to be the number one global cause of death. Despite its mortality, heart failure is still underdiagnosed clinically, especially in a remote area that experiences cardiologists shortage. Existing studies have employed artificial intelligence to help with heart failure screening and diagnosis processes based on echocardiography videos. Specifically, most existing studies use a convolutional neural network that only captures the local context of an image hindering it from learning the global context of an image. Moreover, the frame sampling algorithms only sample certain consecutive frames which makes it questionable whether the dynamic of the left ventricle during a cardiac cycle is included. This study proposed a novel deep learning model consisting of a time-distributed vision transformer stacked with a transformer. The time-distributed vision transformer learns the spatial feature and then feeds the result to the transformer to learn the temporal feature and make the final prediction afterward. We also proposed a frame sampling algorithm by squeezing the video and sampling the frame after a certain interval. Consequently, the video still contains the sequential information up until the end of the video with some in-between frames removed by a certain interval. Thus, the dynamic of the left ventricle is preserved. Our proposed method achieved an F1 score of 95.81%, 96.19%, and 93.43% for the apical four chamber view, apical two chamber view, and parasternal long axis view respectively. The overall trustworthiness of our model is quantified using the NetTrustScore and achieved a score of 0.9712, 0.9767, and 0.9527 for the apical four chamber view, apical two chamber view, and parasternal long axis view respectively.

Fast and accurate post-earthquake building damage assessment is an important task to do to define search and rescue procedures. Many approaches have been proposed to automate this process by using artificial intelligence, some of which use handcrafted features that are considered inefficient. This research proposed end-to-end building damage assessment based on a Siamese neural network. We modify the network by adding a feature concatenation mechanism to enrich the data feature. This concatenation mechanism creates different features based on each output from the convolution block. It concatenates them into a high-dimensional vector so that the feature representation is more likely to be linearly separable, resulting in better discrimination capability than the standard siamese. Our model was evaluated through three experimental scenarios where we performed classification of G1 or G5, G1-G4 or G5, and all the five grades of EMS-98 building damage description. Our models are superior to the standard Siamese neural network and state-of-the-art in this field. Our model obtains f1-scores of 79.47%, 54.09%, 40.64% and accuracy scores of 87.24%, 95.28%, and 42.57% for the first, second, and third experiments, respectively.

One of the major challenges in the change detection (CD) of high-resolution remote sensing images is the high requirement for computational resources. Besides, to get the best change detection result, it must spot only the important changes while omitting unimportant ones, which requires learning complex interactions between multi-scale objects on the images. Despite Convolution Neural Network (CNN) efficiently extracting features from such images, it has a limited receptive field resulting in sub-optimal representation. On the other hand, Vision Transformer (ViT) can capture long-range dependencies. Still, it suffers from quadratic complexity concerning the number of image patches, especially for high-resolution images. Furthermore, both approach can not model the interactions among multi-scale image patches, which is essential for a model to fully understand the natural images. We propose FocalCD, a CD method based on a recently proposed focal modulation architecture capable of learning short and long dependencies to solve this problem. It is attention-free and does not suffer from quadratic complexity. Also, it supports learning multi-scale interaction by adaptively selecting the discriminator regions from multi-scale levels. Besides the efficient yet powerful encoder, FocalCD has an effective multi-scale feature fusion and pyramidal decoder network. FocalCD achieves strong empirical results on various CD datasets, including CDD, LEVIR-CD, and WHU-CD. It reaches F1 scores of 0.9851, 0.952, and 0.9616 on datasets CDD, LEVIR-CD, and WHU-CD outperforming state-of-the-art CD methods while having comparable or even lower computation complexity.

Electrocardiograms (ECG) are vital for diagnosing various cardiac conditions but are often corrupted by noise from multiple sources, which can hinder accurate interpretation. Denoising ECG signals is particularly challenging because noise usually overlaps with the frequency range of the signal of interest. We proposed a convolutional autoencoder with sequential and channel attention (CAE-SCA) to address this issue. Sequential attention (SA) is based on long short-term memory (LSTM), which captures causal-temporal relationships. Meanwhile, channel attention (CA) is used to emphasize important features within channels. SA is applied to the skip connection of each encoder block, and CA is applied after each decoder block. We validated the CAE-SCA using the MIT-BIH and SHDB-AF databases as clean ECG signals, with the MIT-BIH Noise Stress Test Database as the noise source. Experimental results give an average SNR value of 16.187 dB, RMSE of 0.059, and PRD value of 18.529 in the MIT-BIH database. While in the SHDB-AF dataset, the model obtained 15.308 dB of SNR, 0.049 of RMSE, and 19.220 of PRD. These results demonstrate our CAE-SCA outperforms all the state-of-the-art methods across all tested metrics. For efficiency, CAE-SCA achieved competitive results in the metrics of floating-point operations (FLOPs), inference time, and total parameters. This allowed CAE-SCA to be implemented in edge devices as we tested using our custom ECG acquisition circuit. A significance test further confirms a statistically significant improvement in SNR values achieved by the CAE-SCA compared to baseline models, suggesting the CAE-SCA’s potential for advancing ECG processing in healthcare applications.