Histopathology is the cornerstone of oncology diagnosis, while whole-slide images (WSIs) enable the transition to digital, quantitative pathology. Leveraging WSIs to accurately predict therapeutic response is increasingly vital for advancing precision oncology and optimizing clinical workflows. However, existing artificial intelligence models for WSI-based immunotherapy response prediction often yield suboptimal performance, as they frequently fail to fully capture the complex multi-scale morphological features and heterogeneous information inherent in clinical pathology data. To address this challenge, we propose ViGNet, a Visual-Global Relation Fusion Network designed for immunotherapy response prediction directly from WSIs. ViGNet is a multimodal framework that effectively integrates histopathological image features with specific clinical data modalities, including gene expression profiles and cancer type text. The proposed architecture features two primary encoders: a multi-scale visual encoder and a gene-driven encoder. The visual encoder utilizes pre-trained CTransPath and an MSCNNPath to extract pyramidal image features, which are integrated with cancer-type diagnostic priors. Meanwhile, the gene-driven encoder converts gene-guided information into a slide-level global relational prior. This global relational prior was then injected into the patch-level multi-scale visual representation learning via a top-down SRFF module, and the resulting representations were subsequently passed to a classification head to predict immunotherapy response. Across one internal and two independent external cohorts, ViGNet achieves 82.55% discrimination ability for immunotherapy response prediction, outperforming the baseline methods. These results underscore ViGNet's potential as a clinical decision-support tool, aiding personalized treatment planning and enabling informed clinical decision-making through accurate and interpretable predictions.

Accurate identification of lung cancer subtypes is critical for the diagnosis and treatment of pulmonary lesions in clinical practice. Deep learning-based diagnostic systems for surgical lesion section images of lung tumors can assist clinicians in efficiently and accurately recognizing lung cancer subtypes, enabling targeted therapies. However, existing methods often fail to fully leverage prior knowledge from medical images, making them susceptible to interference from high-intensity light regions caused by external illumination. Moreover, many existing methods suffer from high computational complexity and excessive parameters, which hinder their practical application. To address these challenges, this paper proposes MTPKDistillNet, a novel framework designed for subtype recognition of lung cancer from surgical lesion section images. The framework incorporates a prior knowledge module to extract edge and texture features from images. It then performs multi-level information distillation from multiple teacher networks to enhance the feature extraction capability of the student network. Finally, the framework employs decoupled and category relation distillation, allowing the student network to effectively learn both the predicted outputs and the class relevance information from the teacher network. This enhances the diagnostic capability of the model. Experimental results on a clinical lung cancer surgical slide dataset demonstrate that MTPKDistillNet achieves a classification accuracy of 87.55

In recent years, virtual immunohistochemical (IHC) staining, which converts hematoxylin and eosin (H&E) images into IHC images, has emerged as a promising technology in digital histopathology. Most existing methods rely on paired H&E and IHC patches extracted from adjacent tissue sections for supervised training. However, tissue misalignment and tissue loss between adjacent sections lead to inconsistent training pairs, limiting the models' ability to produce accurate staining results. To address this issue, we propose ConCLR, a two-stage virtual IHC staining framework based on context-aware contrastive learning, designed to handle inconsistently paired patches. Our method is built on the assumption that for a given mini-patch in the H&E patch, there may exist a corresponding mini-patch in the reference IHC patch exhibiting a similar Pos/Neg pathological pattern. If such a mini-patch exists, it is typically located spatially close to the H&E mini-patch due to the local consistency of tissue structure. In the first stage, we leverage this assumption to design a similarity-guided mini-patch sampling (SGMS) module. For each mini-patch anchor in the staining results, SGMS searches within the real IHC patch to find the most similar mini-patch to serve as the positive sample for contrastive learning, enabling effective supervision despite mild tissue misalignment. In the second stage, we design a context-aware adaptive refinement module, which addresses significant inconsistencies between training pairs caused by potential tissue loss, by expanding the search range of positive samples to include neighboring patches. Extensive experiments on two network backbones across four virtual IHC staining tasks demonstrate the effectiveness of our ConCLR. Evaluations include qualitative and quantitative assessments of staining results, as well as downstream diagnostic performance. In addition to experiments on existing public datasets, we collected a PanCK-NSCLC dataset by acquiring H&E and pan-cytokeratin staining images from the same lung tissue sections via destaining and restaining. This dataset offers significantly improved tissue alignment compared to those derived from adjacent sections, with the aim of facilitating further progress in virtual IHC staining.

The accurate identification of low-risk and high-risk lung tumors is essential for clinicians to develop lung cancer treatment strategies during surgery. Despite the achievements in deep learning, most medical image analysis applications are still hampered by the difficulty of obtaining large amounts of labeled data. In this paper, a semi-supervised deep learning framework, DS-FixMatch, is proposed for identifying lung tumors intra-operatively while alleviating the issue of sparse annotations. DS-FixMatch combines selective labeling and semi-supervised training: (1) For the acquired unlabeled images, a subset that best represents the distribution of the entire dataset is selected by an unsupervised algorithm. Compared to the traditional random labeling strategy, this method can avoid introducing samples that are highly influenced by the intraoperative environment for labeling. This subset is then sent to a human expert for labeling. (2) Supervised training is performed using labeled images. For the remaining unlabeled samples, DS-FixMatch utilizes model predictions to generate pseudo-labels for consistency regularization, further enhancing the model’s generalization ability. A dataset consisting of 2221 natural images, each capturing the Region of Interest (ROI) in lung tumors, is constructed to evaluate the effectiveness of the designed framework. Experiments show that DS-FixMatch leads in performance for the task of lung tumor recognition compared to other baselines.

Lung cancer is a most prevalent and deadly type of cancer worldwide. In clinical practice, it is crucial to accurately identify the risk level of lung cancer for diagnosing and treating lung lesions. Current AI methods for lung cancer image diagnosis mainly rely on CT images, histopathological images, and a few methods based on surgical lesion section images. However, there is a great need for methods for surgical lesion section images that can help surgeons quickly decide the surgical plan during the operation. Deep learning methods among them often supervise only the output layer of the network, while ignoring the constraints on the intermediate hidden layers, resulting in insufficient accuracy of high-risk and low-risk classification. This paper aims to design a method for the task of classifying surgical lesion section of lung tumours into high-risk and low-risk categories. The EDSNet model based on deep supervision is proposed, which enhances the performance of the hidden layer through deep supervision. Moreover, the MBConv module of baseline model is replaced with the SimAM-MBConv module proposed in this paper, which further enhances the representation of pathological features by deriving 3D attention weights for feature maps within one layer. Several experiments had been conducted on a clinical lung tumour surgical lesion section image dataset, achieving the classification accuracy of EDSNet is 95.90% with an AUC of 0.9935. Compared with most existing methods for similar tasks, our method is optimal in most of the evaluation metrics. The experimental results demonstrate the effectiveness and sophistication of our method for high-risk and low-risk classification of surgical section images of lung tumours, which can serve as a reference for doctors to determine the subsequent surgical plans and treatment strategies.

Pectus excavatum (PE) is one of the most common congenital sternal deformities. Accurate preoperative diagnosis of PE is of great significance for subsequent correction and improvement of the patient's quality of life. However, current diagnostic methods rely on the calculation of some PE indices, which is a heavy workload for physical therapists and suffers from measurement errors. To address this issue, we propose an end-to-end automatic assessment of PE, which features a cascaded structure of CNN and LSTM. Specifically, the high-level feature representations of CT images are extracted by the pretrained CNN and then processed through the LSTM layers for classification. In addition, we build up a medical image dataset for PE diagnosis by collecting chest CT images of 42 subjects. Results on this dataset show that the proposed CNN-LSTM framework achieves a relatively high accuracy of 90.20%, which provides a new perspective for the automatic diagnosis of PE in clinics.

So far, in the study of neurological brain disorder diagnosis, there are mainly two kinds of work that exploits graph via Graph Convolutional Networks (GCN). Though they have achieved remarkable success, neither of them is able to simultaneously account for the brain region level correlation and subject level correlation. To tackle this issue, we design a Graph-In-Graph Convolutional Networks(GIGCN) framework, which turns out to inherit the merits of the two kinds of existing work. Specifically, we propose a graph-in-graph data structure that consists of two types of graph, namely internal brain connective graph and external population graph. In this data structure, an internal graph is assigned to each subject whose nodes represent brain regions and edges incorporate associations between brain regions, while the external graph is a population graph generated from the demographic of all subjects where edges describe the similarity between subjects. After that, our proposed framework employs GCN to explore some deeper relationships for better disease prediction. Results on the ABIDE dataset validate the effectiveness of our proposed method.