Accurate and automated segmentation of gliomas from Magnetic Resonance Imaging (MRI) is crucial for effective diagnosis, treatment planning, and patient monitoring. However, the aggressive nature and morphological complexity of these tumors pose significant challenges that call for advanced segmentation techniques. This review provides a comprehensive analysis of Deep Learning (DL) methods for glioma segmentation, with a specific focus on bridging the gap between research performance and practical clinical deployment. We evaluate over 80 state-of-the-art models published up to 2025, categorizing them into CNN-based, Pure Transformer, and Hybrid CNN-Transformer architectures. The primary objective of this paper is to critically assess these models not only on their segmentation accuracy but also on their computational efficiency and suitability for real-world medical environments by incorporating hardware resource considerations. We present a comparison of model performance on the BraTS datasets benchmark and introduce a suitability analysis for top-performing models based on their robustness, efficiency, and completeness of tumor region delineation. By identifying current trends, limitations, and key trade-offs, this review offers future research directions aimed at optimizing the balance between technical performance and clinical usability to improve diagnostic outcomes for glioma patients.

Gliomas are aggressive brain tumors that require accurate imaging‐based diagnosis, where automated segmentation plays a central role in assessing tumor morphology and guiding treatment decisions. Manual delineation of gliomas is time‐consuming and prone to variability, motivating the use of deep learning to improve consistency and alleviate clinical workload. However, existing methods often fail to fully exploit the information available in multi‐parametric MRI (mp‐MRI), particularly inter‐slice contextual information. In addition, they typically require considerable computational resources and often lack robustness across different tumor types and imaging conditions. We present GBT‐SAM, a parameter‐efficient deep learning framework that adapts the large‐scale Segment Anything Model (SAM) to volumetric mp‐MRI data. GBT‐SAM reduces training complexity by selecting fewer than 2.6% of slices per scan while still incorporating all four MRI modalities, thereby preserving essential tumor‐related information at minimal computational cost. Furthermore, our model is trained using a two‐step fine‐tuning strategy that combines a depth‐aware module to capture inter‐slice correlations with lightweight LoRA‐based adaptation layers, resulting in only 9.97 M trainable parameters, the lowest among existing SAM‐based approaches. GBT‐SAM is trained exclusively on the BraTS Adult Glioma dataset and achieves a Dice score of 92.66. Also, it is evaluated on additional Meningioma, Pediatric Glioma, and Sub‐Saharan Glioma datasets in order to demonstrate its robust generalization performance. These results highlight GBT‐SAM as a computationally efficient and domain‐robust framework for brain tumor segmentation from mp‐MRI. Our code and models are available at https://github.com/vpulab/gbt‐sam .

Access to reliable medical images is essential for clinical training. To address this need, this paper presents an extended version of MIRAGE, a multimodal retrieval and generation system that utilizes a shared latent space to process medical queries by retrieving real images from the ROCO dataset, generating synthetic scans, and providing LLM-based clinical descriptions alongside dual-concept visual comparisons. To overcome previous computational limits and the lack of clinical validation, we introduce three core enhancements: first, an Auto-$\alpha$ module to dynamically weight visual and textual similarities; second, the integration of LCM-LoRA to accelerate synthetic image generation; and third, an automated clinical auditor based on Gemini 2.5 Flash. Experimental results demonstrate that Auto-$\alpha$ improves retrieval accuracy for heterogeneous queries, reaching $38.83$% Top-1 Recall over a 65,419-image gallery and outperforming nine fusion baselines evaluated under a unified configuration, with a controlled ablation attributing most of this gain to learning the weight rather than merely making it query-adaptive, while the LCM-LoRA module reduces computational costs by a factor of 12.5× in CPU environments, with a blinded radiologist evaluation confirming only a small drop in clinical quality. Furthermore, the clinical auditor achieves a 0.805 Pearson correlation against an expert radiologist, effectively correcting the systematic overestimation of traditional CLIP scores. Finally, the optimized platform is publicly deployed on Hugging Face.