Quantitative analysis of biomarker expression is crucial for cancer diagnosis and prognosis. Traditional methods rely on immunohistochemistry (IHC) staining, which can take several days, delaying clinical decisions. To address this issue, this study introduces a novel deep-learning-based method that infers the H-score, a measure of biomarker abundance, directly from readily available hematoxylin and eosin (H&E) stained slides, eliminating the need for IHC staining. We tested our method on H&E slides from breast cancer and renal cell carcinoma, finding an average Pearson correlation coefficient of 0.92 compared to existing IHC-based methods. Additionally, the generated H-score heatmaps for whole slide images stained with H&E demonstrated a strong spatial correlation with the ground truth, with cosine similarity exceeding 0.86 for most biomarker heatmaps. The results indicate that our method can reliably infer the expression of cancer biomarkers from H&E slides alone.Clinical relevance- This approach has the potential to streamline pathology workflows and enhance the efficiency and comprehensiveness of biomarker evaluation in clinical practice.

This study presents a novel machine learning-driven drug repositioning pipeline to identify potential drug candidates for the Classical Subtype of pancreatic ductal adenocarcinoma (PDAC). Our approach integrates the Graph Attention Neural Network version 2 (GATv2) model as the backbone to predict target genes of PDAC, achieving a high area under the curve (AUC) of 0.911. A robust two-step cross-validation workflow was also implemented to enhance model reliability and generalizability. Finally, PharmOmics network analysis was conducted for potential drug candidates' prediction. Through our pipeline, we successfully identified at least five promising drug candidates for PDAC, including Indomethacin, Sorafenib, Valproic Acid, Cyclosporine, and Irinotecan, showcasing the potential of our drug repositioning methodology. These findings demonstrated the effectiveness of machine learning in drug discovery and paved the way for further clinical exploration of these candidates for PDAC treatments.Clinical Relevance- This study leveraged machine learning for drug repositioning with robust modelling design. There were at least five promising drug candidates predicted through our pipeline. Our findings may assist clinicians in considering the proposed drug candidates to enhance treatment strategies and improve patient outcomes in PDAC.

Ensemble Learning is a machine learning method that enhances overall predictive performance by combining multiple base learners. However, most current ensemble learning approaches employ average fusion methods, which overlook the consistency and diversity of individual model predictions and are unable to adaptively handle testing data. This paper introduces an Evidence Reasoning (ER) rule ensemble learning method that unifies model adaptation, uncertainty estimation, and confidence calibration within a single framework, thereby providing a more reliable model to aid physicians in decision-making. We evaluated our approach in predicting locoregional recurrence in Head and Neck Cancer (HNC). Compared to the previously proposed ERE, the ER-rule ensemble model achieved a 4.1% improvement in ACC.Clinical Relevance-This ER-rule ensemble model demonstrates a more reliable approach to predicting locoregional recurrence in head and neck cancer, enabling timely clinical intervention and potentially improving patient outcomes.

Accurate brain tumour segmentation in Multimodal Multi-parametric Magnetic Resonance Imaging is critical for clinical intervention but remains challenging due to complex boundaries and heterogeneous morphology. Existing U-Net architectures often struggle to capture global context and accurately delineate tumours with irregular or diffuse boundaries, limiting their clinical applicability. To mitigate these challenges, We introduce a novel dual-path 3D U-Net framework that integrates Optimal Transport (OT) theory into feature fusion, enabling improved alignment of global and local features. The architecture includes a Global Context Path (via Pyramid Pooling) and a Local Detail Path (with multi-scale convolutions), capturing complementary spatial features. The Optimal Transport Fusion module leverages Optimal Transport theory to efficiently align and merge features from both paths, offering a principled alternative to traditional fusion strategies such as concatenation or attention mechanisms. Additionally, we incorporate an edge-aware loss function based on 3D Sobel operators, which refines boundary precision in segmentation masks. Evaluated on the BraTS 2023 glioma dataset, our model outperforms standard U-Net and other recent state-of-the-art models, demonstrating the effectiveness of Optimal Transport-based feature fusion in enhancing brain tumour segmentation accuracy.

Accurate segmentation of post-operative glioma poses a critical task in medical image processing due to the importance of evaluating therapeutic regimens and guiding the subsequent treatment. High annotation difficulties and costs exacerbate the difficulty of segmentation. Consequently, training an end-to-end segmentation model from scratch using annotated clinical data is infeasible. In addition, although foundation models have made significant progress in medical image segmentation, their direct application to post-operative glioma segmentation still faces challenges. In this paper, we propose a method to simulate post-operative giloma based on pre-operative data resection, and introduce a strategy to fine-tune a foundation model using simulated post-operative data. Post-operative glioma generation strategy combines noise-disturbed polyhedron and level set model. Specifically, the noise-disturbed polyhedron is utilized to simulate the residual cavity, while the level set is employed to mimic the control of resection levels among physicians with varying levels of experience. During the foundation model segmentation phase, a fine-tuning strategy incorporating the gray-level distribution of the tumor region for prompt optimization is taken into account. We evaluated the proposed method on two public post-operative glioma datasets and one private dataset, achieving improvements in the dice coefficient by 15.1%, 8.2%, and 14.2%, respectively, compared to state-of-the-art methods.Clinical Relevance- Our methodology facilitates precise localization of postoperative residual tumors by physicians, offering pivotal insights for the formulation of radiotherapy, chemotherapy, and subsequent follow-up protocols, thereby aiding in clinical decision-making and treatment strategies.

Despite the transformative impact of deep learning on text, audio, and image datasets, its dominance in tabular data, especially in the medical domain where data are often scarce, remains less clear. In this paper, we propose X2Graph, a novel deep learning method that achieves strong performance on small biological tabular datasets. X2Graph leverages external knowledge about the relationships between table columns, such as gene interactions, to convert each sample into a graph structure. This transformation enables the application of standard message passing algorithms for graph modeling. Our X2Graph method demonstrates superior performance compared to existing tree-based and deep learning methods across three cancer subtyping datasets.Clinical relevance- This work advances the application of deep learning solutions to cancer diagnosis, particularly in scenarios where only limited tabular data is available.

The tumor physical microenvironment (TPME), encompassing matrix stiffness, solid stress, hydraulic pressure, and matrix architecture, has been widely revealed to affect the initiation, progression, and metastasis of cancer. However, the underlying mechanisms remain unintelligible. In this work, a triparallel-channel tumor extravasation chip was developed, and the TPME was constructed by integrating the cellular matrix with adjustable mechanical properties. The results indicate that a softer subendothelial matrix led to a higher proportion of tumor cell extravasation across the vascular wall, accompanied by a greater invasion depth, highlighting the critical role of subendothelial matrix mechanical properties in regulating tumor extravasation. This chip provides a novel platform for exploring the impact of TPME on the nonrandom distribution of tumor metastasis sites.Clinical Relevance- The tumor extravasation chip can provide a theoretical foundation for understanding the tumor extravasation mechanism, establish an evaluation model for assessing the efficacy of anticancer drugs in inhibiting tumor metastasis, and aid in formulating personalized treatment plans based on individual tumor characteristics.

Lung cancer is the most common cause of cancer-related death worldwide. The detection of lung nodules from Computed Tomography (CT) scans is essential for assessing disease progression, monitoring treatment response, and guiding therapeutic strategies. Deep learning has emerged as a powerful tool for image segmentation, demonstrating significant potential in medical imaging applications. This work aims to introduce U-Grad, a novel model designed for lung nodule segmentation from 2D CT slices. It integrates an encoder that generates heatmaps using Gradient-weighted Class Activation Mapping (Grad-CAM), which are then concatenated with CT slices and fed into a Reduced U-Net to enhance nodule representation. The Reduced U-Net is characterized by an encoder-decoder structure whose maximum depth, in terms of filter size, is (256,256), Additionally, it employs the Leaky Rectified Linear Unit as an alternative activation function, enhancing its representational capacity. NSCLC Radiogenomics dataset from The Cancer Imaging Archive was used to train and test the proposed U-Grad for 100 epochs. The performance of both the Reduced U-Net and U-Grad models was evaluated using the Dice Coefficient (DC) and the Intersection over Union (IoU) metrics. The results demonstrate that both models outperform existing models in the literature. The Reduced U-Net achieves a DC of 93.15% and an IoU of 89.02%, while U-Grad achieves a DC of 91.27% and an IoU of 86.26% in test set. Although both models exhibit comparable performance, U-Grad demonstrates slightly lower overfitting, making it a more robust alternative. Moreover, U-Grad's ability to generate interpretable heatmaps enhances its utility for clinical applications and research, particularly in resource-limited settings where annotated data are scarce.Clinical relevance- U-Grad is an innovative and effective lung nodule segmentation model that leverages explainable AI techniques to enhance its performance, interpretability and generalizability.

Metastatic colorectal cancer (mCRC) remains a major clinical challenge, with only a subset of patients responding to first-line chemotherapy. Predicting chemoresistance could help personalize treatment and avoid unnecessary toxicity, and AI-based models have shown promise in addressing such predictive tasks. Among the different data sources used in medical AI, histopathological images offer valuable insights, but their variability, particularly in staining, poses a significant challenge. Stain normalization techniques aim to standardize color variations, improving the consistency of AI-driven analysis. Convolutional-based methods are widely used but require selecting a target patch, a process typically performed visually or randomly. The impact of this selection on normalization performance, however, remains unclear. In this study, we investigated how target patch selection influences stain normalization and compared these methods with a generative model approach (CycleGAN), which eliminates the need for a reference patch. Our results showed that different target patches not only altered the color appearance of normalized images but also affected their structural content. While the generative model resolved the target patch selection issue, its overall performance was moderate. Furthermore, stain normalization influenced feature extraction, with the color-deconvolution-based method yielding more relevant features. In the chemotherapy response prediction of mCRC patients, the two stain normalization techniques resulted in different results, with higher performance reached when using the color-deconvolution-based stain normalization method. The best classifier achieved an average cross-validated AUC value of 0.83 and 0.73 on the training and test set, respectively, demonstrating high potential in correctly predicting mCRC response to therapy.Clinical Relevance- In this study, stain normalization influenced both feature predictivity and classifier performance, emphasizing the importance of effectively managing staining variations across patients and medical centers to ensure robust and reliable predictive modeling. Our best model achieved good results in predicting the chemotherapy response of mCRC patients, demonstrating its potential for assisting treatment decisions.

Segmentation of breast ultrasound images is crucial for the early and accurate diagnosis of breast cancer. In this study, we propose TRAM-UNet (Transformer and Region Attention Module-Based U-Net), a novel deep learning model that integrates Transformer blocks and a Region Attention Module (RAM) to improve segmentation performance. TRAM-UNet achieves average Dice scores of 88.56 ±0.91%, 84.68 ±1.20%, and 83.96 ±1.34% on the BUS-BRA, BUSI, and BLUI datasets, respectively, significantly outperforming both U-Net and U-Net + Transformer across all datasets. These results demonstrate TRAM-UNet's ability to refine boundaries, enhance segmentation accuracy, and adapt to different lesion characteristics, underscoring its potential to advance breast ultrasound segmentation and clinical diagnosis.Clinical Relevance- This study is clinically relevant as it demonstrates the potential of deep learning in improving breast ultrasound image segmentation. With further research and optimization, this approach could contribute to more precise and automated breast cancer diagnosis in clinical practice.