Clinical & Experimental Cardiology
Open Access

Significance of Heart Disease based on Deep Hierarchy

Samba Guindo^{* *}Correspondence: Samba Guindo, Department of Cardiology, Cardiology Service of Grand Yoff Teaching Hospital, Dakar, Senegal, Email:

Author info »

Description

One of the most used diagnostic radiology test techniques is the chest X-ray, which helps skilled radiologists identify patients who may be at risk for lung and cardiovascular problems. However, it is still difficult for experienced radiologists to evaluate hundreds of instances in a short amount of time; hence deep learning techniques are being used to solve this issue. The previous classification technique did not function well since the diseases had hierarchical characteristics and relationships with one another. Some GCN-based models are introduced in order to combine the features retrieved from the images to produce predictions in order to extract the correlation features among the diseases. Backbone with high quality image attributes can make this scheme operate nicely.

The cost of computation is crucial to this plan. However, a quick prediction in diagnostic radiology is also necessary, particularly in an emergency or at a location with limited computing resources. To address this need for a quick prediction and high accuracy, we suggested the SGGCN, an effective convolutional neural network with GCN. SGGCN utilizes the Shuffle Ghost Blockbuilt SGNet-101 as its backbone to extract features at a cheap computational cost. A new GCN architecture is created to mix information from many levels together in the GCNM module so that we can leverage various hierarchical characteristics and make the GCN scheme faster at the same time. This allows us to make sufficient use of the information in GCN. The CheXPert datasets experiment provided the proof showing that SGGCN performs admirably. The SGGCN achieves 0.7831 (3.08%) test AUC with parameters 1.2M (73.73%) and FLOPs 3.1B (80.82%) compared to GCN and ResNet-101 backbone (test AUC 0.8080, parameters 4.7M and FLOPs 16.0B), whereas GCN with MobileNet (Sandler and Howard, 2018) backbone achieves 0.7531 (6.79%) test AUC.

Millions of people's lives are at danger for cardiopathy and lung illness, yet the majority of these conditions can be avoided with Chest X-ray (CXR) equipment. CXR technology is now routinely used to check for heart and lung illness, helping with clinical diagnosis and therapy. Convolutional Neural Networks (CNN) and Bayesian models are two algorithms that have been created to analyses and forecast diseases from CXR images, and they have a significant impact. On the one hand, they enable expert radiologists to process a large number of radiological samples while simultaneously reducing the workload of professional radiologists due to the high speed of calculation. On the other hand, some low-risk radiological samples can be removed using these algorithms with a very low false-negative rate, allowing expert radiologists can more easily identify the potentially dangerous materials.

CNN-based models have the ability to extract picture information and employ fully connected layers to forecast the future. The output space is combinatorial, which makes the multi-label task more difficult compared to multi-class picture classification. Since the invention of deep learning, attention has increasingly been paid to Convolutional Neural Networks (CNNs), a type of deep network that is particularly well suited for hierarchical classification. In order to improve the accuracy of the ImageNet classification task, ResNet was proposed as a deep neural network feature extraction method. ResNet is currently widely utilized as a foundation for extracting features, and a pertained model is employed to speed up the training process.

The method in the traditional image classification task might not work since the chest illness recognition challenge is a multi-label classification problem and the label (diseases) contains hierarchical properties. Deep learning has been used in various activities that are crucial for safety and security due to its exceptional performance, including self-driving, malware detection, identification, and anomaly detection.

Researchers have excelled at image classification tasks and made significant progress in the segmentation and classification of medical images thanks to the advent of deep learning. Since the diseases in the chest disease recognition challenge contain hierarchical structures and co-occurrence features, unique strategies should be used to approach this hierarchical multi-label learning classification task. Both the ChestX-ray14 dataset and the CheXpert dataset contain hierarchical multi-label features. Additionally, certain approaches utilizing probability modeling, attention learning, and graph neural networks are also introduced to learn the hierarchical features.

Conclusion

Many attempted to forecast the conditional probability for each label and improved this model with unconditional probability. Guan and Huang created an attention module to acquire normalized attention scores, using ResNet-50 or DenseNet-121 as the backbone, then combined the features from the backbone and the attention scores into a residual attention block to provide classifications. This study proposes the effective X-ray classification algorithm SGGCN, which uses the CheXpert datasets and the SGNet-101 backbone created with the ShuffleGhost Module to classify chest diseases. We also compare ResNet-101 with GCN and MobileNetV2 with GCN's AUC, trainable parameters, and FLOPs. It is discovered that despite a large drop in the trainable parameters and FLOPs, SGGCN still maintains a high AUC on the validation and testing set.