The detection of new Multiple Sclerosis (MS) lesions refers to identifying newly appeared or progressing lesion areas within specific time periods from MS-related regions in longitudinal medical images. In recent years, many deep learning methods have been proposed for this field. However, existing methods often fail to fully utilize prior knowledge about MS lesions to suppress interference from non-lesion signals. In this paper, we propose the Lesion-Guided New Lesion Detection Network (LGNLDNet) for enhancing the detection of new MS lesions. Specifically, The model first learns MS lesion-related prior knowledge, enabling effective extraction of lesion-specific features during the global encoding of images. Since globally encoded image features inherently contain a large amount of non-lesion signals, lesion features are prone to being overshadowed by these non-lesion signals when the model analyzes new lesions. Therefore, the model further extracts prior knowledge represented by local lesion features. This prior knowledge is used to construct prompt information. The proposed Lesion-Guided Differential Enhancement Feature Fusion Module (LG-DEFF) then employs the prompt to suppress irrelevant background interference, thus effectively capturing discriminative features of newly emerging MS lesions. Additionally, we introduce the Multi-Level Feature Fusion Module (MLFF), which is cascaded in a pyramidal structure to effectively fuse complementary information from feature maps at different levels. Experimental results on the MICCAI-21 dataset demonstrate that the proposed method outperforms state-of-the-art approaches.

Self-supervised low-dose computed tomography (LDCT) denoising remains challenging due to the difficulty of recovering fine texture details while suppressing structured artifacts. To address these limitations, we propose the Semantic-Guided Artifact-Aware Diffusion model (SGAA). SGAA is designed with a two-stage denoising framework. In the first stage, it leverages multi-scale semantic features of NDCT extracted from a unified latent space to guide the reverse process, thereby mitigating prediction errors caused by domain shift and achieving the removal of random noise in LDCT. In the second stage, the artifact-aware mechanism identifies artifacts by utilizing the attenuation characteristics of unknown noise in the iterative reverse process, so as to eliminate residual artifacts from the first stage. Experiments on the Mayo LDCT dataset demonstrate that SGAA outperforms state-of-the-art methods in both quantitative metrics and visual fidelity when only using NDCT image data.

In clinical practice, MR images with incomplete modalities lead to severe degradation of tumor segmentation performance. We propose an incomplete modality Segmentation network based on shared latent feature compensation (SLFCNet) to address the challenge of robust segmentation in the presence of arbitrary missing modalities. Specifically, SLFCNet includes a set of hybrid encoders, a shared encoder, and a Multimodal correlation Modeling Module (MCMM). The hybrid encoder employs the mamaba mechanism to extract modality-specific features; the shared encoder extracts shared latent features. The MCMM introduces the Transformer to realize effective information fusion of the two types of features. The fused features have richer semantic information and can produce more accurate segmentation results. Comprehensive experiments on BraTS2018 and BraTS2021 datasets verify the effectiveness of our method.

Radiology report generation aims to provide comprehensive clinical descriptions and ease radiologists' workloads. Previous research has explored using knowledge graphs and auxiliary classification tasks to enhance the model's ability to generate accurate reports. However, due to the lack of information in the knowledge graphs or insufficient class label information, these methods fail to provide models with clinical severity information about the same disease at different stages of development, resulting in less accurate reports. To address this issue, we propose a Severity-Guided Radiology Report Generation method (SR2Gen), which guides the model in identifying internal severity variations of the disease from both explicit and implicit dimensions. Specifically, SR2Gen includes two innovative modules: a Knowledge Enhancement Module (KEM) and a Disease Severity-Aware Module (DSAM). First, KEM explicitly guides the report generation model by constructing a knowledge graph containing disease severity information as prior knowledge. Secondly, DSAM enhances the severity-aware classifier using pseudo-labels generated through momentum distillation and further incorporates an adaptive disease severity learning method, implicitly guiding the model to learn disease progression. Extensive experiments and analyses on IU X-Ray and MIMIC-CXR datasets demonstrate that SR2Gen outperforms previous state-of-the-art methods.

In video analysis, existing approaches achieve excellent performance under preset conditions, but the scarcity of training videos limits deep learning development. To address this, this paper introduces two external resources to alleviate the video shortage: image data and cross-institutional data. We propose a novel method, IOLCIA, which combines image-only learning and cross-institutional learning to maximize the use of image resources. Additionally, it leverages the differences and similarities in data from various institutions to enhance feature learning. Comparison experiments show that our method outperforms models trained solely on images, with only a slight decrease in effectiveness compared to those trained on videos.