Background:Accurate and interpretable brain tumor classification from MRI images remains a key challenge in medical image analysis, particularly when using publicly available datasets of moderate size. Objectives:This study investigates the performance of a ConvNeXt-Tiny based framework for four-class brain tumor classification glioma, meningioma, pituitary tumor, and no tumor and compares it with established convolutional architectures. Materials and methods:Using transfer learning and identical experimental settings, ConvNeXt-Tiny was evaluated against DenseNet169, Xception, MobileNetV3-Large, CNN+DenseNet169, and ResNet50. Standard evaluation metrics (accuracy, precision, recall, and F1-score) were used, and Grad-CAM was applied to visualize model attention for interpretability. Generalization was further assessed using an independent dataset. Results:ConvNeXt-Tiny achieved high overall performance (accuracy = 0.9924, F1-score = 0.9918), comparable to DenseNet169 and Xception but with lower computational cost. The model maintained stable learning behavior, minimal overfitting, and consistent accuracy on unseen data. Grad-CAM visualizations confirmed that the network focused on clinically relevant tumor regions, improving transparency and reliability of predictions. Conclusion:ConvNeXt-Tiny provides a strong and efficient baseline for interpretable brain tumor classification, balancing accuracy and computational efficiency. While the results are promising, differences among recent architectures were modest, and clinical validation using multi-center MRI datasets is necessary to confirm broader applicability.

Neuroimaging correlates of cognitive stages in Parkinson's disease: A multimodal MRI study.

Sturge-Weber syndrome (SWS) is a rare neurocutaneous disorder associated with venous capillary malformations, atrophy, and calcifications. Longitudinal imaging is limited by risks of sedation and gadolinium exposure in children. To evaluate whether strategically acquired gradient echo (STAGE), a rapid multi-contrast quantitative MRI method, can reliably detect vascular and parenchymal abnormalities in SWS compared with conventional pre-/post-contrast MRI. Observational cross-sectional. Twenty-two patients with unilateral SWS diagnosed by previous MRI (13 female; ages 2-24 years). 3T/T1-weighted (T1W) and T2-weighted (T2W) turbo-spin-echo, fluid attenuated inversion recovery, and a 3D gradient echo-based STAGE sequence providing T1, proton density (PD), T2*, and R2* maps, susceptibility-weighted imaging (SWI), quantitative susceptibility mapping (QSM), T1W with enhanced gray matter to white matter contrast (T1WE), and synthetic images of T2W, FLAIR, and gradient echo images. Conventional MRI and STAGE images were reviewed in 10 patients (training group), side-by-side, to determine the STAGE-derived images that identify SWS abnormalities, including leptomeningeal venous capillary malformations (LVCM), enlarged deep medullary veins, choroid plexus enlargement, cerebral atrophy, and calcifications. In the remaining test group of 12 patients, three reviewers scored these abnormalities on STAGE images and compared them with scores from conventional MRI. Interrater reliability with intraclass correlation coefficient (ICC), Spearman's rank correlation, Wilcoxon signed-ranked test, Mann-Whitney U-test, Fisher's exact test. Statistical significance level was set as p < 0.05. LVCMs were visualized on STAGE with SWI and R2*. Calcifications were differentiated from venous abnormalities using PD, T1WE, synthetic gradient echo, and QSM. STAGE-derived scores had excellent interrater reliability (ICCs > 0.90) and were similar to the conventional MRI scores despite some minor differences in some individual cases (total scores from conventional MRI vs. STAGE 8.9 vs. 8.7, p = 0.29). STAGE provided rapid, non-contrast, multi-parametric imaging that reliably detected vascular and parenchymal SWS abnormalities seen on conventional MRI. 2. Stage 3.

The value of fluorine-18 fluorodeoxyglucose positron emission tomography/magnetic resonance imaging (18F-FDG PET/MRI) in characterising pleural effusion and detecting pleural metastasis for staging lung cancer.

This study investigates the development of a nomogram for predicting the short-term collapse progression of osteonecrosis of the femoral head by integrating clinical data with radiomics features obtained from hip joint MRI. The study involved 364 patients with osteonecrosis of the femoral head who had not yet severe collapsed(Collapse < 2mm or no collapse) from two medical centers, selected from a major medical center. MRI images of their hip joints were analyzed to extract radiomics features. A clinical model was developed using criteria such as ARCO classification, CJFH classification, JIC classification, and the modified Kerboul angle. The outcome variable was the occurrence of collapse progression within one year post-examination. Features most significantly associated with collapse progression were identified from both radiomics and clinical data, which were then integrated into separate models. A combined Nomogram model was constructed by merging the clinical and radiomics models. The performance of these models was compared to assess their effectiveness in predicting collapse progression. The Nomogram model demonstrated superior predictive performance compared to both the clinical and radiomics models across all cohorts. In the external validation set, the Nomogram achieved an AUC of 0.919 and an accuracy of 0.879, outperforming the clinical (AUC=0.809) and radiomics (AUC=0.875) models. Statistical significance was confirmed between the Nomogram and clinical model in all cohorts. In summary, our study demonstrates that a nomogram combining hip MRI-based radiomics with clinical data shows superior predictive performance compared to clinical-only models for assessing short-term collapse risk in osteonecrosis of the femoral head.

This paper presents a novel multi-stage dementia diagnosis framework integrating a Swin Transformer architecture with explainable AI for brain MRI analysis. The proposed approach addresses two critical challenges: capturing both local and global structural features through hierarchical Vision Transformer processing, and providing clinically interpretable decisions via Grad-CAM visualization.Our model was evaluated on a Kaggle dataset comprising 6,400 MRI images across four dementia stages: non-demented (3,200), very mild (2,240), mild (896), and moderate (64). The dataset was split into 70% training, 15% validation, and 15% testing. Experimental results demonstrate superior performance with 97.3% accuracy, precision ranging from 94.8-100%, recall between 91.1-100%, and a macro F1-score of 96.5%. Statistical validation through 5-fold cross-validation (96.8% ± 0.4%) confirms robustness.The SwinGrad-CAM component successfully identifies clinically relevant biomarkers, including hippocampal atrophy and ventricular enlargement, aligning with established neurological indicators. For very mild cases, heatmaps highlight early temporal lobe changes, while moderate cases show intense activation in regions with severe cortical atrophy. This interpretable AI framework offers a robust solution for early intervention, precise staging, and personalized treatment planning in dementia care, enabling clinicians to make informed decisions through visual validation of model reasoning while bridging the gap between deep learning performance and clinical trust.

The clinical diagnosis of major depressive disorder (MDD) has long relied on subjective assessments, underscoring the need for objective and quantifiable automated diagnostic methods. Existing neuroimaging-based multimodal fusion methods face three core challenges: the lack of mechanisms that dynamically assess modality reliability, the absence of effective prediction uncertainty quantification, and site effects in multi-center data that constrain model generalizability. To mitigate these issues, we propose an uncertainty-aware evidential multimodal fusion network for large-scale multi-center MDD classification. We design deep functional feature extractors (DFFE) and dual-stream hierarchical feature fusion (DS-HFF) modules to encode functional magnetic resonance imaging (fMRI) and structural MRI (sMRI) features, respectively. An Evidential Multimodal Fusion (EMF) strategy based on Dempster-Shafer theory (DST) is designed to convert each modality's categorical support into subjective opinions with explicit uncertainty. Reliable cross-modal evidence integration is achieved through context-adaptive discounting and conflict-aware combination. Site-adversarial regularization is incorporated to learn site-invariant pathological feature representations. Evaluated on the REST-meta-MDD dataset comprising 1,601 subjects from 16 sites under a leave-one-site-out cross-validation (LOSO-CV) scheme, the proposed method outperforms recently published methods. Uncertainty analysis demonstrates superior calibration over Softmax and MC Dropout baselines, and the uncertainty-based rejection mechanism further improves the classification performance. Interpretability analysis identifies key brain regions highly consistent with established MDD pathophysiological findings, further validating the clinical reliability of the framework. Experimental results indicate that our proposed method has the potential to provide a robust, trustworthy, and interpretable solution for multi-center MDD diagnosis.

MRI is important for cardiac disease evaluation, but accurate diagnosis remains challenging in less experienced centers. Although large language models (LLMs) have shown promise in medical imaging diagnosis, their application in cardiac MRI is limited. LLMs may be effective in achieving cardiac MRI diagnosis based on standardized descriptions. Retrospective. A total of 203 hypertrophic cardiomyopathy, 186 dilated cardiomyopathy, 46 hypertensive heart disease, 198 ischemic cardiomyopathy, 38 constrictive pericarditis, 45 cardiac amyloidosis, 91 myocarditis, and 144 normal controls. Balanced steady-state free-precession, short tau inversion recovery, and breath-hold inversion-recovery segmented gradient-echo sequences at 3.0 T. Clinical and cardiac MRI information from each subject was converted into standardized descriptions and input into Generative Pre-trained Transformer-4.5 (GPT-4.5), GPT-4 Omni (GPT-4o), Deepseek-V3, and Deepseek-R1 LLMs. Cardiac MRI information included LV function, wall thickness and motion, and abnormalities on T2WI, perfusion and late gadolinium enhancement sequences. Each model was asked to generate an imaging diagnosis. In addition, a medical student (8 months experience) and three radiologists (junior, mid-level and senior: with 3, 6, and 10 years' experience, respectively) provided diagnoses based on cardiac MRI images and clinical information. Frequency-weighted sensitivity and specificity were calculated. The diagnostic performances of the LLMs and human readers were compared using the McNemar test with Bonferroni correction. A p value < 0.05 was considered significant. All LLMs showed excellent frequency-weighted specificity (0.973-0.983). The frequency-weighted sensitivities of all LLMs were not significantly different from that of the junior radiologist, were significantly higher than that of the medical student, and significantly inferior to those of the senior radiologist (GPT-4.5: 0.863, GPT-4o: 0.821, Deepseek-V3: 0.843, and Deepseek-R1: 0.851 vs. junior radiologist: 0.850, all adjusted p = 1.000; vs. medical student: 0.731, all adjusted p < 0.001; vs. senior radiologist: 0.942, all adjusted p < 0.001). Additionally, the mid-level radiologist achieved a frequency-weighted sensitivity of 0.895, outperforming all LLMs except GPT-4.5. LLMs may generate accurate diagnoses from standardized cardiac MRI descriptions, potentially benefiting less experienced physicians. Stage 5.

Alzheimer's disease (AD) emerges from multi-scale interactions between molecular pathology and disruptions in large-scale brain network dynamics. Understanding how these processes co-evolve and relate to disease stages is essential for advancing complex systems models of aging and AD, and for developing system-informed interventions. However, progress has been limited by a lack of large-scale longitudinal data. To address this, we examined the longitudinal relationship between subsystems of the default mode network (DMN) (posterior DMN, ventral DMN, anterior dorsal DMN) using task-free functional MRI (fMRI) and amyloid positron emission tomography (PET) imaging in a large longitudinal cohort spanning the clinico-biological spectrum of AD (n = 1,451; 2,763 time points) using mixed-effect models. We also assessed whether patterns of DMN connectivity predicted conversion to amyloid positivity, mild cognitive impairment (MCI), and dementia using Cox proportional hazards models. Our findings reveal a dynamic interplay between amyloid accumulation and connectivity within and between DMN subsystems, with both hyper- and hypoconnectivity emerging across DMN subsystems in association with increasing amyloid burden. Importantly, survival models showed that DMN connectivity patterns predicted conversion to critical stages of the disease, including not only conversion to MCI and dementia, but also conversion to amyloid positivity in otherwise clinically unimpaired individuals who were amyloid negative at baseline. These associations were independent of age, APOE4 status, sex, education, and in-scanner motion. These results support a model in which breakdowns in tightly regulated feedback loops governing DMN physiology represent a core systems-level pathophysiology of AD. Notably, this functional dyshomeostasis precedes detectable amyloidosis on imaging. Future studies should focus on the development of robust biomarkers of brain function that can be applied at the individual level, which could in turn help support the development of therapeutic approaches targeting system-level pathophysiology.

Cardiac vasculitis represents a heterogeneous group of immune-mediated disorders that can involve the coronary vessels, myocardium, valvular apparatus and pericardial tissues. Despite its rarity, cardiac vasculitis may result in significant clinical sequelae such as acute coronary syndrome, heart failure, cardiac arrhythmias and pericarditis. Diagnosis is challenging because symptoms are often non-specific and overlap with other cardiovascular conditions. Early recognition is therefore crucial to prevent delayed treatment and disease progression. Advances in non-invasive multimodality imaging and collaborative cardio-rheumatology care have transformed recognition and management of this disease spectrum. Emerging techniques such as hybrid positron emission tomography-cardiac MRI and quantitative CT imaging permit in-vivo characterisation of inflammation. As per European Alliance of Associations for Rheumatology recommendations, treatment requires early intensive immunosuppression to induce remission, coupled with comprehensive cardiovascular risk management. Additional research is required to validate imaging-guided management algorithms, refine vasculitis-specific cardiovascular risk and define long-term outcomes across disease subtypes.

Automatic generation of clinical reports from medical images is a relevant task capable of reducing the workload of radiologists and standardizing documentation. In this paper, we investigate an approach to generating structured reports from brain MRI data using a pre-trained multimodal SigLIP2 model as a feature extractor. We propose an architecture in which visual embeddings obtained from a frozen SigLIP2 are projected into the representation space of the GPT-2 language model for subsequent text generation. Experiments were conducted on the open-access BIOSE MRI dataset, containing 34 pairs of "MRI image + clinical report". It is shown that the proposed approach generates semantically meaningful reports, achieving quality comparable to more complex architectures with substantially lower computational costs. Additionally, the influence of pre-training SigLIP2 on a classification task (Brain3-Anomaly-SigLIP2 version) on generation quality is investigated. The results demonstrate the potential of using frozen vision encoders in medical generative tasks under data-scarce conditions.

360-Degree Virtual Reality Consultations for Patients Undergoing Spine Surgery: A Single Center Pilot Study.

IntroductionAccurate preoperative measurement of pancreatic cystic lesions is critical for surgical decision-making, particularly around the 3cm threshold frequently cited in resection guidelines. However, imaging modalities may provide discrepant size estimates, potentially altering management.MethodsWe retrospectively reviewed patients from a single endoscopist practice (November 2011-November 2024) who underwent EUS-FNA for pancreatic cystic lesions with available MRI and/or CT imaging results. For each modality, the largest recorded dimension was extracted. One-way ANOVA was performed to assess mean size differences, and concordance across the 3cm surgical threshold was calculated with Cohen's κ.ResultsA total of 190 patients with 246 patient encounters were included. Mean cyst size was 2.93 ± 1.32cm on MRI (n = 147), 2.95 ± 1.53cm on EUS (n = 246), and 3.20 ± 1.73cm on CT (n = 159). The differences were not statistically significant (P = .194). Across modalities, the mean absolute difference in size ranged from 0.7 to 0.9cm, with outliers exceeding 2cm. Concordance at the 3cm cutoff was common; there was good agreement between EUS and MRI (κ = 0.643; P < .001) and moderate agreement between EUS and CT (κ = 0.472; P < .001).ConclusionSubstantial variability exists in preoperative size assessment of pancreatic cystic lesions, with 7.5-20% of cysts misclassified around the 3cm cutoff. Given that cyst size is a key determinant of surgical intervention, reliance on a single modality may introduce bias into clinical decision-making. Multimodality assessment reduces discordance and may improve risk stratification.

Small-molecule inhibitors targeting isocitrate dehydrogenase (IDH) 1/2-mutant proteins have demonstrated benefit in IDH1/2-mutant gliomas. However, responses assessed by conventional MRI measurements are infrequent, delayed, and difficult to interpret, highlighting the need for early biomarkers of treatment benefit. In this study, we investigated 3,4-dihydroxy-6-[18F]-fluoro-L-phenylalanine positron emission tomography (18F-DOPA-PET) and MRI responses in patients with IDH1/2-mutant glioma receiving IDH inhibitors (IDHi). Patients with IDH1/2-mutant glioma receiving IDHi as part of trials or expanded access programs with pre- and posttreatment MRI and 18F-DOPA-PET were included. Evaluations included 2D/3D measurements on T2-weighted fluid-attenuated inversion recovery images; T1-post-contrast, perfusion, and diffusion imaging for MRI; and metabolic tumor volume (MTV), total lesion glycolysis, and tumor-to-background ratios (TBR) for 18F-DOPA-PET. Disease response evaluation using volumetric assessments, RANO 2.0, and PET RANO 1.0 criteria were compared and correlated with outcomes. From 2021 to 2025, 20 patients with IDH1/2-mutant glioma (8 with astrocytoma and 12 with oligodendroglioma) receiving IDHi (4 receiving ivosidenib and 16 receiving vorasidenib) were analyzed. Significant reductions in 18F-DOPA-PET parameters including TBRmean, TBRmax, and MTV were observed in 10 of 20 patients, aligning with observed changes in perfusion and diffusion imaging. Nine partial responses and one complete response were identified using 18F-DOPA-PET, whereas both volumetric and standard 2D morphologic MRI assessments indicated stable disease as best response. PET response on MTV was correlated with prolonged tumor control. These results highlight the potential of 18F-DOPA-PET and advanced MRI sequences as valuable complements to standard RANO 2.0 MRI evaluations for assessing treatment response in patients with glioma undergoing IDHi therapy.

Stroke survivors often face long-term cognitive and motor deficits. Brain age gap (BAG), the difference between chronological age and age estimated based on MRI data, has emerged as a biomarker for neurodegeneration. While prior work links BAG to stroke outcomes, the relationship between BAG and cerebral microbleeds (CMBs), particularly infratentorial CMBs common in hypertensive arteriopathy, remains unclear. The sensorimotor network (SMN) is highly susceptible to both direct and remote injury after stroke and is structurally and functionally interconnected with infratentorial regions via pathways such as the corticospinal tract. Vascular disruption in the cerebellum or brainstem may therefore have downstream effects on supratentorial SMN regions, making this network a biologically relevant target for investigating BAG-CMB relationships. We analyzed data from 1725 stroke patients in the Stroke Outcomes Optimization Projects. Two trained raters manually counted infratentorial CMBs on susceptibility-weighted MRI images (SWI), while BAG was computed using the automated volBrain BrainStructureAges pipeline on T1-weighted images. Spearman correlations tested associations between CMB count and regional BAG in 14 a priori brain regions of interest (ROI) and results were conditioned for age, sex, race, white matter hyperintensities, hypertension, type of scanner, and total ischemic lesion volume. Infratentorial CMB count was positively correlated with BAG in 9/14 sensorimotor regions: right precentral gyrus medial segment (r (213) = 0.186, p = 0.007), left precentral gyrus medial segment (r (213) = 0.186, p = 0.007), right postcentral gyrus medial segment (r (213) = 0.202, p = 0.004), right postcentral gyrus (r (213) = 0.202, p = 0.004), left postcentral gyrus (r (213) = 0.161, p = 0.021), right parietal operculum (r (213) = 0.198, p = 0.004), right central operculum (r (213) = 0.195, p = 0.005), right precentral gyrus (r (213) = 0.184, p = 0.008), and left postcentral gyrus medial segment (r (213) = 0.192, p = 0.006). Our findings suggest that infratentorial microvascular injury is associated with accelerated aging in functionally connected motor cortices. This supports a network-level model of stroke-related brain aging, with implications for predicting sensorimotor outcomes. BAG may serve as a sensitive marker for cerebrovascular injury and guide targeted rehabilitation efforts.

Differential diagnosis of migraine and tension-type headache based on brain volumes using machine learning methods.

Primary CNS T-cell lymphoma (PCNSTL) is an extremely rare and aggressive form of non-Hodgkin lymphoma. Diagnosis is often challenging because of the nonspecific clinical presentation, which can lead to delays in treatment. This study aims to analyze the clinical, histopathologic, and neuroimaging characteristics of PCNSTL. In this retrospective, multicentric cohort study, histologically confirmed PCNSTL cases without evidence of systemic disease were selected from pathology databases at 3 academic neuro-oncology centers (University Hospital of Munich, University Hospital of Heidelberg, Massachusetts General Hospital in Boston) during the period from 2008 to 2024. Retrospective data, including demographics, histopathology, immunophenotyping, multimodal MRI, and PET imaging, treatment lines, and outcome data were extracted from medical records and analyzed. We evaluated 16 patients (11 male) with a median age of 50 years (19-76 years) and a median Karnofsky performance status of 80% (20%-100%). T-cell clonality was confirmed in 9/15 (60%) tested patients with typical T-cell receptor gene rearrangement patterns. MRI scans at primary diagnosis showed predominantly supratentorial parenchymal lesions with prominent contrast enhancement in 14/16 (87.5%) and in more than one-third multifocal lesions (7/16, 43.75%). High intratumoral susceptibility signal was frequently observed (7/16, 43.75%). [18F]fluorodeoxyglucose-PET and [18F]fluoroethyltyrosine-PET imaging revealed only low to moderate tracer uptake in 7/8 examined patients (87.5%). Median progression-free survival was 4 months, and median overall survival was 97.5 months. Although treatment protocols varied, the use of methotrexate (MTX)-based chemotherapy combined with autologous stem-cell transplant (ASCT) was associated with most favorable outcome (95% CI 2-87, p < 0.0275). In 1 case of ALK1-positive PCNSTL, persistent complete remission was achieved after treatment with the ALK-inhibitor lorlatinib. Although PCNSTL is exceptionally rare, we identified distinct neuroimaging patterns showing highly aggressive features on MRI but hypometabolic PET imaging, which may assist in identifying future PCNSTL cases. Despite the limited cohort size, our findings suggest that MTX-based chemotherapy with ASCT may translate into favorable outcome. We report on an ALK1-positive PCNSTL case with sustained complete remission after targeted therapy with lorlatinib.

To characterise aetiologies and prenatal neuroimaging findings in fetuses with intracranial haemorrhage (ICH), and evaluate the ability of imaging to suggest underlying causes. We retrospectively reviewed fetuses with ICH diagnosed prenatally at a single tertiary fetal centre between 2015 and 2025. Cases with a known mechanism at presentation were excluded. Ultrasound (US) and MRI images were reassessed by fetal medicine and neuroradiology specialists blinded to aetiology. Clinical, laboratory, genetic, autopsy and placental findings were integrated to determine underlying causes and assess imaging patterns. Sixty-seven fetuses were included; a definitive aetiology was identified in 38.8% (26/67). Genetic aetiologies were detected in 27.7% (13/47) of cases with advanced genetic testing, most commonly associated with cerebral small vessel disease (CSVD). Other causes included infection (5/67, 7.5%), fetal anaemia (6/67, 9%), and coagulopathy (3/67, 4.5%, of which one was genetic). Twenty-one cases (31.3%) remained unexplained despite comprehensive investigation, and 20 (29.9%) had incomplete investigations. Intraventricular haemorrhage (IVH) was present in 91% (61/67), with periventricular haemorrhagic infarction (PVHI) in 55.2% (37/67). CSVD was associated with multifocal PVHI, porencephaly, white matter injury and frequent brainstem involvement. In fetal anaemia, cerebellar haemorrhage was common, with brainstem and vermian involvement predominantly in non-parvovirus cases. Intraparenchymal haemorrhage without IVH occurred mainly with coagulopathy or infection. Although some phenotypic clustering was observed, there was substantial overlap among aetiologic groups. Genetic disorders, particularly CSVD, account for a significant proportion of prenatal ICH, supporting routine genomic testing. Imaging can suggest probable causes, but overlap limits its diagnostic specificity.

ABSTRACT Brain tumors are among the leading causes of cancer-related mortality worldwide, making early and accurate diagnosis essential for improving patient survival and treatment outcomes. Manual interpretation of MRI scans by radiologists remains challenging due to human fatigue, inter-observer variability, and complex tumor appearances, with reported diagnostic error rates of 20 to 30 percent. This paper proposes an explainable deep learning-based brain tumor classification system using ResNet50 with a three-phase progressive fine-tuning strategy, trained on 7,210 MRI images across four classes namely Glioma, Meningioma, No Tumor, and Pituitary tumor. Three explainable AI methods namely Grad-CAM, Grad-CAM++, and HiRes-CAM were comparatively evaluated, and HiRes-CAM was selected based on its superior sharpness scores for precise tumor localization. The system additionally extracts ten clinically relevant spatial parameters including anatomical region, tumor volume, hemisphere, midline shift, and edema volume to support medical decision-making. The proposed system achieved a test accuracy of 95.00%, sensitivity of 0.950, specificity of 0.985, and F1-score of 0.9488, with AUC scores reaching 0.9997 for Pituitary tumor. The complete system is deployed as a web-based clinical application named NeuroScan AI, supporting real-time analysis at 85 to 120 milliseconds, batch processing, patient history management, and automated PDF report generation. Key Words : Brain Tumor Classification, Deep Learning, ResNet50, HiRes-CAM, Explainable AI, MRI Analysis, Spatial Parameter Extraction, Transfer Learning, Clinical Decision Support.

Brain cancer remains a critical global health challenge, where early and accurate diagnosis remains a critical challenge in clinical practice. Current supervised learning methods for tumor classification face substantial limitations due to their dependence on large labeled datasets requiring costly pixel-level annotations, susceptibility to annotation biases, and poor generalization across diverse populations. To address these challenges, this paper proposes Retrospection Dropout Bare-Bones Particle Swarm Optimization (RDBPSO), a novel feature-based classification framework that requires only image-level class labels without the need for pixel-level annotation or manual segmentation masks. The proposed RDBPSO introduces two key innovations: (1) a retrospection mechanism that maintains dual-layer memory structures (optimal and sub-optimal solutions) to enhance particle diversity and prevent premature convergence, and (2) a dropout strategy that reduces computational complexity through intelligent particle interaction sampling. Extensive experiments on an 800-image brain MRI dataset demonstrate RDBPSO’s superior performance. The proposed method achieves 90.12% classification accuracy, outperforming standard PSO (89.25%), GMM (77.50%), and K-means (72.75%), while delivering robust clustering quality with an ARI of 0.6436, NMI of 0.5511, and FMI of 0.8229. These results demonstrate the algorithmic promise of RDBPSO as an annotation-efficient framework for brain tumor MRI classification, warranting further investigation on more diverse and clinically representative datasets.