Diabetic retinopathy (DR) is a very common complication of diabetes that can lead to retinal damage, affecting vision. If not detected and diagnosed in time, it may result in blindness. With the development of deep learning, various automatic diagnosis models for DR have been proposed. Most models are based on convolutional neural networks, but due to the small size of convolution kernels in shallow networks, the receptive field is limited, preventing the capture of global information. This limitation forces models to require more convolutional layers to expand the receptive field and capture global information related to DR. In this paper, we use a combined module of multiscale convolution and convolutional self-attention, and we adjust the module structure, allowing the model to have both long-range dependency modeling capability and CNN’s multiscale local feature extraction capability, while having fewer model parameters than the selected comparison models. Our proposed MSAmix-Net achieved an accuracy of 82.3% in the five-class classification task on the combined dataset of APTOS-2019 and Messidor-2, and an accuracy of 93.9% in the three-class classification task on the DRAC-2022 dataset. The experiments demonstrate that the model can effectively capture and represent the complex features of diabetic retinopathy.

The characteristics of the lesion areas in medical images are complex, and existing fully connected layer-based neural networks still cannot address the issue of linear kernels, making them inadequate for handling the nonlinear classification problem of medical image data. This paper investigates a multi-scale feature joint prediction Kolmogorov-Arnold Network convolutional network combined with a spatial attention mechanism for medical image classification applications. In this article, we propose the Multi-Scale Feature prediction network combined with Kolmogorov-Arnold Network. In this model, convolutional blocks are concatenated to output feature maps of different scales. These feature maps are then passed through a spatial attention module, and the output is weighted and summed using learnable weights in the Kolmogorov-Arnold Network layer for classification. The model is trained, tuned, and tested on three publicly available medical image datasets, including the skin cancer image dataset, COVID-19 lung CT dataset, and brain tumor MRI dataset. Experimental results show that the proposed model demonstrates strong performance in classification tasks across four datasets, achieving an accuracy of 87.27% for binary skin cancer classification, 94.12% for the three-class COVID-19 CT classification, and 97.48% for the four-class brain tumor CT classification. This performance outperforms the classification accuracy of the selected comparison methods. Additionally, we applied a perturbation rate of 0.5% and 1.0% using the FGSM method on the test sets of the three datasets and re-tested, finding that our method achieved higher accuracy than the comparison methods. Comprehensive experimental results validate the effectiveness of our proposed method, demonstrating its applicability and robustness in different medical scenarios. This approach is expected to enhance the performance of medical image-assisted classification.

Abdominal CT images are important clues for diagnosing liver cancer lesions. However, liver cancer presents challenges such as significant differences in tumor size, shape, and location, which can affect segmentation accuracy. To address these challenges, we propose an end-to-end 3D segmentation algorithm, RMCNet. In the shallow encoding part of RMCNet, we incorporated a 3D multiscale convolution (3D-Multiscale Convolution) module to more effectively extract tumors of varying sizes. Moreover, the convolutional block attention module (CBAM) is used in the encoding part to help the model focus on both the shape and location of tumors. Additionally, a residual path is introduced in each encoding layer to further enrich the extracted feature maps. Our method achieved DSC scores of 76.56% and 72.96%, JCC scores of 75.82% and 71.25%, HD values of 11.07 mm and 17.06 mm, and ASD values of 2.54 mm and 10.51 mm on the MICCAI 2017 Liver Tumor Segmentation public dataset and the 3Dircadb-01 public dataset, respectively. Compared to other methods, RMCNet demonstrates superior segmentation performance and stronger generalization capability.

Diabetic Retinopathy (DR) is a common complication of diabetes that can lead to vision loss if not detected early. Deep learning models, particularly Convolutional Neural Networks (CNN), have shown impressive results in DR detection. However, these models are susceptible to adversarial attacks, such as the Fast Gradient Sign Method (FGSM), leading to misclassification of detection data. In order to improve the robustness of the Diabetic Retinopathy (DR) detection model against FGSM attacks, we propose a new neural network in this paper that combines the Radial Basis Function (RBF) network with the modern CNN ResNet. We first trained ResNet-101 and the proposed ResNet-101-RBF on the aptos2019-blindness-detection dataset. After training, we evaluated the performance of the two models on normal DR data, and next, we evaluated the performance of the two models on adversarial data generated by FGSM attacks. Our results show that the accuracy of ResNet-101 in normal data is 78.26%, and the accuracy of ResNet-101-RBF is 75.26%, with the accuracy of ResNet-101 being 3% higher than that of ResNet-101-RBF in normal data. However, on adversarial samples generated with perturbation rates of 0.01, 0.05, and 0.1, the accuracy of ResNet-101-RBF is 6%, 5.5%, and 4.75% higher than that of ResNet-101, respectively. The ResNet-101-RBF model can maintain a high detection accuracy in normal data while having higher robustness against FGSM attacks compared to the ResNet-101 model.