Unsupervised anomaly detection methods primarily focus on modeling the distribution of normal samples at image/feature level. Significant deviation from the modeled distribution is then considered as anomaly. Yet, each normal sample may have its own unique characteristic drifting from the idea distribution, making it difficult to distinguish between anomaly and characteristic. In this paper, we propose a memory-based Coarse-to-Fine Feature Fusion (C3F) module to tackle this challenge. Specifically, we construct a group of memory banks that model the feature distribution of normal samples at various levels of granularity. The memory-based C3F is applied to each skip-connection between the encoder and the decoder, and progressively removes anomaly while maintaining characteristic. This helps to reconstruct defect-free image with characteristic preserved, encouraging large (respsmall) deviation from the modeled distribution for anomaly (respcharacteristic). Besides, we also introduce a novel rough anomaly score map Guided Segmentation (GS) module to achieve precise anomaly localization. Extensive experiments on widely used VisA and MVTec-AD benchmarks demonstrate the wide-ranging applicability of the proposed method termed C3FGS on industrial components of various forms. The implementation code is publicly available athttps://github.com/LZL501/c3f_industrial_anomaly_detection.

Often, the characteristics of thyroid nodules need to be determined by fine needle aspiration (FNA) biopsy. The increasing applications of machine learning and deep learning algorithms provide alternative noninvasive methods to study thyroid nodules on ultrasound images. Many studies examined the feasibility of convolutional neural networks or radiomics feature extraction to analyze the characteristics of thyroid nodules. In this study, we built an automated radiomics analysis system by combining thyroid segmentation via U-Net and radiomics feature extraction. Our proposed machine learning-based automated radiomics analysis was compared to a deep learning-based convolutional neural network method in a two-center thyroid nodule classification task. It is shown that the automated radiomics analysis can accurately segment thyroid nodules to facilitate clinical diagnosis by achieving dice scores of 0.77 and 0.74 on internal and external sets respectively. In addition, the proposed automated radiomics analysis approach can improve sensitivity, negative predictive value (NPV) and positive predictive value (PPV) by 41.2%, 3.5% and 7.5% respectively, while reducing the false negative rate by 41.1%.

Deep learning algorithms can evaluate large and complex sets of data, offering various support for medical imaging analysis. Previous works have explored applications of deep learning to measure leg lengths more efficiently. These previous studies provide evidence to suggest deep-learning algorithms can improve efficiency with high levels of accuracy and speed. In this retrospective study, we utilize deep learning-based convolutional neural networks, programmed with input from a human expert, to identify key points and measure leg length. We collected frontal computed tomography (CT) scout radiographs from pre-operative CT scans of patients undergoing evaluation for knee arthroplasty from diverse sources to both train and test the model. We prepared a DenseNet121 model to predict and identify key points, which were then used to develop patch-based models. We applied separable convolutional layers to complete the analysis. The data reflects that 1) separable convolution exhibits lower mean absolute error (MAE) and increased convergence speed as compared to global average pooling layers and 2) optimal learning rates, batch size, and patch size can be achieved to present the least MAE. Our findings provide useful information and an automated tool to assist radiologists to diagnose leg length discrepancy in clinical practice.