Amir Hossein
Pouria

This paper introduces SVD-GAT-GB, a hybrid framework designed to conquer the pervasive issue of network sparsity in graph link prediction. By combining truncated Singular Value Decomposition (SVD) for structural graph densification, Graph Attention Networks (GAT) for context-aware node embeddings, and Gradient Boosting (GB) for final prediction, the model extracts deep structural features that traditional methods miss. Extensive benchmarking across diverse datasets demonstrates a massive boost in performance, outperforming existing baselines by approximately 32% in F1-score and 11% in AUC while maintaining exceptional robustness across varying network types.

Targeting the severe shortage of expert annotations in medical AI, this multi-center study establishes a highly data-efficient semi-supervised learning (SSL) pipeline for CT-based lung cancer survival prediction. Leveraging a combination of 1,218 PyRadiomics features and a pseudo-labeling framework tested across 12 independent cohorts, the top-performing architecture (Random Forest feature selection paired with XGBoost) outperformed standard supervised alternatives by up to 17%, securing an external accuracy of 0.88. Crucially, the system maintains stable, low-variance predictions even when restricted to a mere 10% of labeled data, while integrated SHAP analysis uncovers the specific radiomic signatures driving long-term survival predictions to ensure clinical transparency.

This work presents PySERA, a modular, object-oriented Python library engineered to fix the reproducibility and deep learning integration gaps in modern radiomics pipelines. Re-implementing the popular MATLAB-based SERA framework, it extracts 557 standardized handcrafted features alongside high-dimensional deep learning embeddings from pre-trained architectures (ResNet, DenseNet, VGG) via a parallelized, multi-core engine. Validation across 8 public datasets proved that PySERA achieves >94% compliance with the Image Biomarker Standardization Initiative (IBSI)—substantially outperforming PyRadiomics in feature consistency—while delivering superior downstream classification accuracies (0.43–0.84) and full compatibility with PyTorch and MONAI.

To demystify the "black box" of medical AI, this study introduces BM1.0, a dual-dictionary framework that translates abstract radiomic math into standard BI-RADS clinical vocabulary used by breast radiologists. Evaluated on dynamic contrast-enhanced MRI data from 1,549 patients to differentiate triple-negative breast cancer (TNBC) from other subtypes, the framework paired an Extra Trees classifier with SHAP analysis to achieve a 0.83 cross-validation accuracy. Instead of outputting uninterpretable numerical arrays, the framework maps predictive features directly to observable clinical traits—proving, for instance, how high mathematical 'Sphericity' correlates to round tumor margins and how low 'Busyness' reflects homogeneous internal enhancement—thereby fostering deep clinical trust.

Fusing multi-sequence MRI protocols (T1, T2, T1CE, FLAIR) with semi-supervised pseudo-labeling, this study offers a highly reliable, non-invasive alternative to surgical brain biopsies for identifying glioblastoma IDH mutation status. Utilizing data from 1,329 patients across 8 clinical sites, the optimized pipeline (RFE + SVM) achieved an exceptional cross-validation accuracy of 0.93 and an external validation accuracy of 0.75, visibly outstripping single-sequence approaches. The architecture's core strength lies in its low sensitivity to cohort size, retaining strong performance in data-scarce conditions, while SHAP explainability confirms that semi-supervised learning fundamentally sharpens the discriminative visibility of first-order statistics and wavelet-based biomarkers.

This comprehensive survey maps out the rapid evolution of Graph-LLM integration, capturing how researchers are marrying the structural precision of Graph Neural Networks with the rich contextual understanding of Large Language Models. The paper establishes a multi-perspective taxonomy that categorizes recent breakthroughs across seven core operational axes, including prompt engineering, semantic enrichment, architectural design, and scalability constraints. By analyzing hybrid neuro-symbolic pipelines that blend structured topologies with unstructured natural language reasoning, this work dissects modern evaluation standards and highlights the open technical boundaries preventing fully explainable, open-world graph intelligence systems. The literature is actively tracked via an accompanying open-source repository.

Grounded in the Knowledge-to-Action framework, this project develops a collaborative, human-centered deep learning pipeline to automate lung cancer CT segmentation without sacrificing clinical agency. Benchmarking five state-of-the-art architectures against expert manual contours from 999 patients, a VNet model combined with semi-supervised learning yielded the highest spatial accuracy (Dice 0.83) and exceptional radiomic stability via PySERA metrics. Beyond quantitative success, a qualitative evaluation by six independent physicians showed that while the AI generated smoother, highly reliable peritumoral boundaries, clinicians explicitly favor this technology as an initial interactive masking assistant for manual refinement rather than an autonomous replacement—offering a highly practical roadmap for medical AI adoption.