Deep model with Siamese network for viable and necrotic tumor regions assessment in osteosarcoma

Abstract

To achieve automatic classification of viable and necrotic tumor regions in osteosarcoma, most of the existing deep learning methods can only design a simple model to prevent overfitting on small datasets, which leads to the weak ability of extracting image features and low accuracy of the models. In order to solve the above problem, a deep model with Siamese network (DS-Net) was designed in this paper. Methods: The DS-Net constructed on the basis of full convolutional networks is composed of an auxiliary supervision network (ASN) and a classification network. The construction of the ASN based on the Siamese network aims to solve the problem of a small training set (the main bottleneck of deep learning in medical images). It uses paired data as the input and updates the network through combined labels. The classification network uses the features extracted by the ASN to perform accurate classification. Results: Pathological diagnosis is the most accurate method to identify osteosarcoma. However, due to intraclass variation and interclass similarity, it is challenging for pathologists to accurately identify osteosarcoma. Through the experiments on hematoxylin and eosin (H&E)-stained osteosarcoma histology slides, the DS-Net we constructed can achieve an average accuracy of 95.1%. Compared with existing methods, the DS-Net performs best in the test dataset. Conclusions: The DS-Net we constructed can not only effectively realize the histological classification of osteosarcoma, but also be applicable to many other medical image classification tasks affected by small datasets.

Densely connected attention network for diagnosing COVID-19 based on chest CT

“…From Table 4 , we can see that the classification effect of the DenseANet is significantly better than that of the other four models, indicating that the densely connected attention mechanism is effective. From Table 4 , we can also see that the classification effect of Baseline is better than ResNet101, which is consistent with the conclusion of our previous work [ 40 ]. From the results of the Baseline, Baseline+ARL and DenseANet models, it can be seen that as the number of attention features in the model increases, the better the classification effect of the mode.…”

“…3 (b) into ResNet101 to construct a ResNet101+ARL model for comparing. At the same time, we combined ARL with the backbone network (Baseline) we proposed in the literature [ 40 ] to construct a Baseline+ARL model. In the experiments, we used ResNet101, ResNet101+ARL, Baseline, Baseline+ARL and DenseANet five networks to conduct five-fold cross-validations.…”

“…In this paper, based on our previous work [ 40 ], firstly, a densely connected attention block (DAB) was proposed. Secondly, we replaced the convolutional layer in U-Net with the DAB to construct an Attention U-Net segmentation model (A-U-Net) to segment lung fields in chest CT images.…”

“…3 (d) is the feature extraction block proposed in our previous work [ 40 ]. It was verified in the literature [ 40 ] that the ability of Fig. 3 (d) to extract features is better than that of Fig.…”

“…In order to avoid losing the image characteristics of the lung lesions of COVID-19 patients, we only segmented the lung fields, and used the segmented lung CT images as the input of DenseANet. Inspired by the design idea of DenseNet [ 39 ], on the basis of our previous work [ 40 ], we generated a series of attention features in the construction of the deep model, and densely connected these attention features to enhance the feature extraction capabilities of the model. In summary, our contributions have the following three aspects: We proposed a densely connected attention block (DAB) as the basic feature extraction block of CNN, this novel densely connected form of attention features can enhance the model’s ability to extract features of different pneumonia types, thereby improving the model’s diagnostic effect.…”