Search

The primary prerequisite for multi-lesion segmentation is the simultaneous detection of multiple lesions. Numerous techniques based on the frameworks of simultaneous multi-lesion segmentation and repetitive single-lesion segmentation are continually refining their models to meet this requirement. While simultaneous multi-lesion segmentation techniques may fully exploit interactions between distinct lesions, single-lesion segmentation methods can concentrate only on a particular lesion. However, since different lesions exhibit distinct patterns, it is challenging for these sophisticated models to work properly when several lesions are present simultaneously. We propose a retinal multi-lesion segmentation method by reinforcing single-lesion guidance with multi-view learning. To the best of our knowledge, this is the first work to formulate the retinal multi-lesion segmentation task as a multi-view task. In the multi-view method, each segmentation branch incorporates context guidance of the particular lesion at the input to focus more attention on specific lesions while producing multi-lesion outcomes. For implementing the multi-view method, we design a two-level hierarchical heterogeneous network, whose core element is the multi-view segmentation branches at the second level. The first level provides the context guidance as multi-view cues. Concretely, our network incorporates two complementary models (i.e., U-Net and TransU-Net). Compared with U-Net focusing on local receptive fields, TransU-Net augments U-Net by Transformer, which is adept at establishing long-distance dependence through global attention. The entire network consists of TransU-Net at the first level and a subsequent multi-view U-Net group. It is denoted as Trans2U-Net. Extensive experiments on three datasets demonstrate the effectiveness of the proposed method. • This study treats the retinal multi-lesion segmentation task as a multi-view task. • Each branch of multi-lesion segmentation pays more attention to a particular lesion. • A two-level hierarchical heterogeneous network is designed for context guidance. • Experiments on three public datasets verify the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

As Deep Neural Networks (DNNs) are increasingly deployed in safety-critical scenarios, there is a growing need to address bit-flip errors occurring in hardware, such as memory. These errors can lead to changes in DNN weights, potentially degrading the performance of deployed models and causing catastrophic consequences. Existing methods improve DNNs' fault tolerance or robustness by modifying network size, structure, or inference and training processes. Unfortunately, these methods often enhance robustness at the expense of clean accuracy and introduce additional overhead during inference. To address these issues, we propose S harpness- A ware Minimization for enhancing DNNs' R obustness against bit-flip errors (SAR), which aims to leverage the intrinsic robustness of DNNs. We begin with a comprehensive investigation of DNNs under bit-flip errors, yielding insightful observations regarding the intensity and occurrence of such errors. Based on these insights, we identify that Sharpness-Aware Minimization (SAM) has the potential to enhance DNN robustness. We further analyze this potential through the relationship between SAM formulation and our observations, building a robustness-enhancing framework based on SAM. Experimental validation across various models and datasets demonstrates that SAR can effectively improve DNN robustness against bit-flip errors without sacrificing clean accuracy or introducing additional inference costs, making it a "double-win" method compared to existing approaches. • The first work to uncover the DNNs' intrinsic robustness against bit-flip errors. • The first work to adopt Sharpness-Aware Minimization to resist bit-flip errors. • A valuable and lightweight framework for security-critical scenarios. [ABSTRACT FROM AUTHOR]

Mainstream few-shot segmentation methods meet performance bottleneck due to the data scarcity of novel classes with insufficient intra-class variations, which results in a biased model primarily favoring the base classes. Fortunately, owing to the evolution of the Internet, an extensive repository of unlabeled images has become accessible from diverse sources such as search engines and publicly available datasets. However, such unlabeled images are not a free lunch. There are noisy inter-class and intra-class samples causing severe feature bias and performance degradation. Therefore, we propose a semi-supervised few-shot segmentation framework named F4S , which incorporates a ranking algorithm designed to eliminate noisy samples and select superior pseudo-labeled images, thereby fostering the improvement of few-shot segmentation within a semi-supervised paradigm. The proposed F4S framework can not only enrich the intra-class variations of novel classes during the test phase, but also enhance meta-learning of the network during the training phase. Furthermore, it can be readily implemented with ease on any off-the-shelf few-shot segmentation methods. Additionally, based on a Structural Causal Model (SCM), we further theoretically explain why the proposed method can solve the noise problem: the severe noise effects are removed by cutting off the backdoor path between pseudo labels and noisy support images via causal intervention. On PASCAL-5 i and COCO-20 i datasets, we show that the proposed F4S can boost various popular few-shot segmentation methods to new state-of-the-art performances. [Display omitted] • Incorporating semi-supervised learning into few-shot task to tackle data scarcity. • A ranking algorithm to identify and remove noisy samples in pseudo labels. • Explaining the algorithm using a Structural Causal Model to reduce confounding bias. [ABSTRACT FROM AUTHOR]

Depth completion is a crucial task in autonomous driving, aiming to convert a sparse depth map into a dense depth prediction. The sparse depth map serves as a partial reference for the actual depth, and the fusion of RGB images is frequently employed to augment the completion process owing to its inherent richness in semantic information. Image-guided depth completion confronts three principal challenges: (1) the effective fusion of the two modalities; (2) the enhancement of depth information recovery; and (3) the realization of real-time predictive capabilities requisite for practical autonomous driving scenarios. In response to these challenges, we propose a concise but high-performing network, named CHNet, to achieve high-performance depth completion with an elegant and straightforward architecture. Firstly, we use a fast guidance module to fuse the two sensor features, harnessing abundant auxiliary information derived from the color space. Unlike the prevalent complex guidance modules, our approach adopts an intuitive and cost-effective strategy. In addition, we find and analyze the optimization inconsistency problem for observed and unobserved positions. To mitigate this challenge, we introduce a decoupled depth prediction head, tailored to better discern and predict depth values for both valid and invalid positions, incurring minimal additional inference time. Capitalizing on the dual-encoder and single-decoder architecture, the simplicity of CHNet facilitates an optimal balance between accuracy and computational efficiency. In benchmark evaluations on the KITTI depth completion dataset, CHNet demonstrates competitive performance metrics and inference speeds relative to contemporary state-of-the-art methodologies. To assess the generalizability of our approach, we extend our evaluations to the indoor NYUv2 dataset, where CHNet continues to yield impressive outcomes. The code of this work will be available at https://github.com/lmomoy/CHNet. [ABSTRACT FROM AUTHOR]

As an architectural CNN accelerator integrated into NVIDIA's GPUs, existing research mainly focuses on improving the performance of Tensor Cores. However, the highly integrated Tensor Cores are vulnerable to transient faults (i.e., soft errors), causing catastrophic consequences in safety-critical applications like automatic driving. Thus, it is imperative to estimate the reliability of CNNs on Tensor Cores. However, obtaining a statistically significant resilience profile of CNNs on Tensor Cores with the existing fault injection (FI)-based reliability estimation methods is expensive. To this end, we build TC-SEPM to predict the error resilience of CNNs on Tensor Cores instead of FI methods. To ensure the accuracy of TC-SEPM, we first investigate resilience-related features from program and microarchitecture perspectives. Then, leveraging these heuristic features, we train machine learning models to learn the hidden relationship between error resilience and the investigated features, enabling us to predict the impact of soft errors in Tensor Cores on CNN output. Experimental results show that TC-SEPM achieves high accuracy for individual soft error resiliency prediction and overall program resilience estimation while its overhead is only 1/27 of FI methods. Additionally, TC-SEPM can provide valuable insights for programmers or architects to design more robust CNN models on Tensor Cores. • The first work to predict the soft error resilience of CNNs on Tensor Cores. • The first work to make a hardware–software co-exploration on reliability-relevant features. • A valuable tool for programmers or architects due to its high efficiency and accuracy on soft error reliability evaluation. [ABSTRACT FROM AUTHOR]

Pneumonia can be difficult to diagnose since its symptoms are too variable, and the radiographic signs are often very similar to those seen in other illnesses such as a cold or influenza. Deep neural networks have shown promising performance in automated pneumonia diagnosis using chest X-ray radiography, allowing mass screening and early intervention to reduce the severe cases and death toll. However, they usually require many well-labelled chest X-ray images for training to achieve high diagnostic accuracy. To reduce the need for training data and annotation resources, we propose a novel method called Contrastive Domain Adaptation with Consistency Match (CDACM). It transfers the knowledge from different but relevant datasets to the unlabelled small-size target dataset and improves the semantic quality of the learnt representations. Specifically, we design a conditional domain adversarial network to exploit discriminative information conveyed in the predictions to mitigate the domain gap between the source and target datasets. Furthermore, due to the small scale of the target dataset, we construct a feature cloud for each target sample and leverage contrastive learning to extract more discriminative features. Lastly, we propose adaptive feature cloud expansion to push the decision boundary to a low-density area. Unlike most existing transfer learning methods that aim only to mitigate the domain gap, our method instead simultaneously considers the domain gap and the data deficiency problem of the target dataset. The conditional domain adaptation and the feature cloud generation of our method are learning jointly to extract discriminative features in an end-to-end manner. Besides, the adaptive feature cloud expansion improves the model's generalisation ability in the target domain. Extensive experiments on pneumonia and COVID-19 diagnosis tasks demonstrate that our method outperforms several state-of-the-art unsupervised domain adaptation approaches, which verifies the effectiveness of CDACM for automated pneumonia diagnosis using chest X-ray imaging. • A novel domain adaptation method for small-scale target dataset. • A contrastive learning-based strategy to solve data deficiency problem. • An adaptive feature cloud expansion mechanism to improve model's generalisation ability. • A new SOTA result for automated pneumonia and COVID-19 diagnosis. [ABSTRACT FROM AUTHOR]