For solving the limitations of the current self knowledge distillation including never fully utilizing the knowledge of shallow exits and neglecting the impact of auxiliary exits' structure on the performance of network, a novel self knowledge distillation framework via virtual teacher-students mutual learning named LOTH is proposed in this paper. A knowledgeable virtual teacher is constructed from the rich feature maps of each exit to help the learning of each exit. Meanwhile, the logit knowledges of each exit are incorporated to guide the learning of the virtual teacher. They learn mutually through the well-designed loss in LOTH. Moreover, two kinds of auxiliary building blocks are designed to balance the efficiency and effectiveness of network. Extensive experiments with diverse backbones on CIFAR-100 and Tiny-ImageNet validate the effectiveness of LOTH, which realizes superior performance with less resource by the comparison with the state-of-the-art distillation methods. The code of LOTH is available on Github https://github.com/cloak-s/LOTH.

The Convolutional Neural Networks and Transformer cannot provide satisfactory performance in medical image classification due to insufficient data, high resolution, and a lot of redundancy. To achieve better performance, this paper proposes a structure-fusion network that combines the architecture of convolution and transformer. To reduce the computational overhead incurred by the transformer structure, we optimize it by aggregating adjacent features. We further modify the Multilayer Perceptron using convolution to increase the network capacity. The network is verified on the grading of the Anterior Cruciate Ligament from knee MRI and the screening of pneumonia and COVID-19 from Chest X-ray. Compared with current advanced methods, the proposed network not only reduces FLOPs but also achieves improvement on AUC and F1score.

Abstract Medical image segmentation from noisy labels is an important task since obtaining high‐quality annotations is extremely difficult and expensive. There are a lot of approaches proposed for such task. However, some issues like the overfitting on noisy annotations, the limited learning of boundary features, and no consideration of the corrupted local pixels are still not solved. Therefore, a novel approach named uncertainty‐aware iterative learning (UaIL) is proposed for medical image segmentation with noisy labels. UaIL iteratively and jointly trains two deep networks using the original images and their argumented ones through a joint loss function including softened label loss, hard label loss and consistency loss, which encourages UaIL to produce segmentations that are robust to the perturbations in arbitrary semantic space. The uncertainty of labels is estimated based on the predictions in iterative learning, then the original labels are refined, which improves the learning of boundary features in segmentation. To avoid overfitting, a stopping strategy is designed based on the dice coefficient in iterative learning. Experiments on two public datasets verify the effectiveness of UaIL under different levels of annotation noise. Especially, when there are serious noises in the labels, the dice achieved by UaIL is 1.43% to 15.03% higher than the competing approaches on the two public datasets. The UaIL is further verified on a private dataset, which shows its ability of applying in the real application with noisy labels.

Recently, tensor decomposition approaches are used to compress deep convolutional neural networks (CNN) for getting a faster CNN with fewer parameters. However, there are two problems of tensor decomposition based CNN compression approaches, one is that they usually decompose CNN layer by layer, ignoring the correlation between layers, the other is that training and compressing a CNN is separated, easily leading to local optimum of ranks. In this paper, Learning Tucker Compression (LTC) is proposed. It gets the best tucker ranks by jointly optimizing of CNN's loss function and Tucker's cost function, which means that training and compressing is carried out at the same time. It can directly optimize the CNN without decomposing the whole network layer by layer and can directly fine-tune the whole network without using fixed parameters. LTC is verified on two public datasets. Experiments show that LTC can make a network like ResNet, VGG faster with nearly the same classification accuracy, which surpasses current tensor decomposition approaches.

Automated anomaly detection and localization in musculoskeletal radiographs are essential for large-scale screening in the radiograph workflow. However, the anomaly is a localized pattern that may be affected by the extra irrelevant areas, and the ambiguous and subtle features of abnormal regions are difficult to detect. To tackle these two problems, we propose a deep guided context-aware network (DR-Net) for anomaly detection in musculoskeletal X-rays. Specifically, we design a positional guide module, which embeds the spatial positional information as prior knowledge to guide the network to enhance the feature representation of a specific region. Then, to detect subtle anomalies, we construct a contextual relation module. It can obtain context-aware features by capturing the spatial dependence of any two positions from the entire X-ray image. It combines context appearance information and selects more distinguishable features from space and channel, producing a detailed visualization of the anomaly region. Note that only image-level labels are required. The extensive experiments on the two radiograph datasets show that DR-Net has a promising performance in anomaly detection and localization.

Chronic kidney disease (CKD) is a widespread renal disease throughout the world. Once it develops to the advanced stage, serious complications and high risk of death will follow. Hence, early screening is crucial for the treatment of CKD. Since ultrasonography has no side effects and enables radiologists to dynamically observe the morphology and pathological features of the kidney, it is commonly used for kidney examination. In this study, we propose a novel convolutional neural network (CNN) framework named the texture branch network to screen CKD based on ultrasound images. This introduces a texture branch into a typical CNN to extract and optimize texture features. The model can automatically generate texture features and deep features from input images, and use the fused information as the basis of classification. Furthermore, we train the base part of the network by means of transfer learning, and conduct experiments on a dataset with 226 ultrasound images. Experimental results demonstrate the effectiveness of the proposed approach, achieving an accuracy of 96.01% and a sensitivity of 99.44%.

Abstract In this paper, we propose a real‐time method for rendering soft shadows and inter‐reflections of dynamic objects under complex illumination. In previous methods, many efforts were taken to acquire occlusion and reflection informations for dynamic scene on the fly, and the result image cannot be generated in real time. In our approach, these informations for each object are pre‐computed and stored in well‐defined Spherical Radiance Transport Maps (SRTMs). For distant complex illumination such as environment illumination and area light source, we decompose the illumination to several hundred directional lights. In rendering, we search in SRTMs for occlusion info which may cause shadows, and reflection info which may cause inter‐reflections. Finally we produce realistic soft shadows and inter‐reflections efficiently. Our method is related with but different from previous Pre‐computed Radiance Transfer techniques which are only suitable for static scene. Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: shading and shadowing