CAC: Confidence-Aware Co-Training for Weakly Supervised Crack Segmentation.

Automatic crack segmentation plays an essential role in maintaining the structural health of buildings and infrastructure. Despite the success in fully supervised crack segmentation, the costly pixel-level annotation restricts its application, leading to increased exploration in weakly supervised crack segmentation (WSCS). However, WSCS methods inevitably bring in noisy pseudo-labels, which results in large fluctuations. To address this problem, we propose a novel confidence-aware co-training (CAC) framework for WSCS. This framework aims to iteratively refine pseudo-labels, facilitating the learning of a more robust segmentation model. Specifically, a co-training mechanism is designed and constructs two collaborative networks to learn uncertain crack pixels, from easy to hard. Moreover, the dynamic division strategy is designed to divide the pseudo-labels based on the crack confidence score. Among them, the high-confidence pseudo-labels are utilized to optimize the initialization parameters for the collaborative network, while low-confidence pseudo-labels enrich the diversity of crack samples. Extensive experiments conducted on the Crack500, DeepCrack, and CFD datasets demonstrate that the proposed CAC significantly outperforms other WSCS methods.

Confidence-Aware Severity Assessment of Lung Disease from Chest X-Rays Using Deep Neural Network on a Multi-Reader Dataset.

In this study, we present a method based on Monte Carlo Dropout (MCD) as Bayesian neural network (BNN) approximation for confidence-aware severity classification of lung diseases in COVID-19 patients using chest X-rays (CXRs). Trained and tested on 1208 CXRs from Hospital 1 in the USA, the model categorizes severity into four levels (i.e., normal, mild, moderate, and severe) based on lung consolidation and opacity. Severity labels, determined by the median consensus of five radiologists, serve as the reference standard. The model's performance is internally validated against evaluations from an additional radiologist and two residents that were excluded from the median. The performance of the model is further evaluated on additional internal and external datasets comprising 2200 CXRs from the same hospital and 1300 CXRs from Hospital 2 in South Korea. The model achieves an average area under the curve (AUC) of 0.94 ± 0.01 across all classes in the primary dataset, surpassing human readers in each severity class and achieves a higher Kendall correlation coefficient (KCC) of 0.80 ± 0.03. The performance of the model is consistent across varied datasets, highlighting its generalization. A key aspect of the model is its predictive uncertainty (PU), which is inversely related to the level of agreement among radiologists, particularly in mild and moderate cases. The study concludes that the model outperforms human readers in severity assessment and maintains consistent accuracy across diverse datasets. Its ability to provide confidence measures in predictions is pivotal for potential clinical use, underscoring the BNN's role in enhancing diagnostic precision in lung disease analysis through CXR.

Non-equivalent images and pixels: Confidence-aware resampling with meta-learning mixup for polyp segmentation.

Automatic segmentation of polyp regions in endoscope images is essential for the early diagnosis and surgical planning of colorectal cancer. Recently, deep learning-based approaches have achieved remarkable progress for polyp segmentation, but they are at the expense of laborious large-scale pixel-wise annotations. In addition, these models treat samples equally, which may cause unstable training due to polyp variability. To address these issues, we propose a novel Meta-Learning Mixup (MLMix) data augmentation method and a Confidence-Aware Resampling (CAR) strategy for polyp segmentation. MLMix adaptively learns the interpolation policy for mixup data in a data-driven way, thereby transferring the original soft mixup label to a reliable hard label and enriching the limited training dataset. Considering the difficulty of polyp image variability in segmentation, the CAR strategy is proposed to progressively select relatively confident images and pixels to facilitate the representation ability of model and ensure the stability of the training procedure. Moreover, the CAR strategy leverages class distribution prior knowledge and assigns different penalty coefficients for polyp and normal classes to rebalance the selected data distribution. The effectiveness of the proposed MLMix data augmentation method and CAR strategy is demonstrated through comprehensive experiments, and our proposed model achieves state-of-the-art performance with 87.450% dice on the EndoScene test set and 86.453% dice on the wireless capsule endoscopy (WCE) polyp dataset.

TSegNet: An efficient and accurate tooth segmentation network on 3D dental model.

Automatic and accurate segmentation of dental models is a fundamental task in computer-aided dentistry. Previous methods can achieve satisfactory segmentation results on normal dental models; however, they fail to robustly handle challenging clinical cases such as dental models with missing, crowding, or misaligned teeth before orthodontic treatments. In this paper, we propose a novel end-to-end learning-based method, called TSegNet, for robust and efficient tooth segmentation on 3D scanned point cloud data of dental models. Our algorithm detects all the teeth using a distance-aware tooth centroid voting scheme in the first stage, which ensures the accurate localization of tooth objects even with irregular positions on abnormal dental models. Then, a confidence-aware cascade segmentation module in the second stage is designed to segment each individual tooth and resolve ambiguities caused by aforementioned challenging cases. We evaluated our method on a large-scale real-world dataset consisting of dental models scanned before or after orthodontic treatments. Extensive evaluations, ablation studies and comparisons demonstrate that our method can generate accurate tooth labels robustly in various challenging cases and significantly outperforms state-of-the-art approaches by 6.5% of Dice Coefficient, 3.0% of F1 score in term of accuracy, while achieving 20 times speedup of computational time.