Deep learning-based automated segmentation of neck abscesses on multiparametric MRI.

Automated deep learning-based segmentation and volumetric analysis of meningiomas.

INTRODUCTION Meningiomas are the most common primary intracranial tumors and are frequently monitored over extended periods. Volumetric assessment typically requires manual segmentation, which is time-consuming and associated with interrater variability. This study aimed to develop and validate a deep learning-based model for the automated segmentation of meningiomas and associated peritumoral edema on preoperative magnetic resonance imaging (MRI). METHODS We trained a standard nnU-Net deep learning model on contrast-enhanced T1-weighted and FLAIR MRI scans from 100 patients treated at the University Hospital of Zurich. The model was then externally validated on 88 cases from the meningioma SEG-Class dataset from the Cancer Imaging Archive. Segmentation performance was assessed using the Dice similarity coefficient, Jaccard index, and 95th percentile Hausdorff distance. RESULTS The model achieved mean Dice scores of 0.87 ± 0.23 for meningioma segmentation and 0.63 ± 0.38 for peritumoral edema in internal cross-validation. On the external validation set, the model achieved scores of 0.86 ± 0.17 for meningioma segmentation and 0.31 ± 0.35 for edema. CONCLUSION The deep learning model demonstrated high accuracy in segmenting meningiomas and modest performance for peritumoral edema. These results support the potential utility of automated segmentation tools in clinical workflows. Future work should focus on validating model performance across larger multi-center datasets.

Zero-shot segmentation and tumor diameter measurement of soft tissue sarcomas on MRI using a multimodal large language model.

Deep learning segmentation with uncertainty quantification for spinal tuberculosis on fat-suppressed T2-weighted MRI

Background Spinal tuberculosis is the most common extrapulmonary manifestation of tuberculosis. Magnetic resonance imaging (MRI), particularly fat-suppressed T2-weighted imaging (FS-T2WI), is the modality of choice for preoperative evaluation; yet indistinct lesion boundaries render manual delineation subjective and poorly reproducible. Although deep learning segmentation has advanced considerably, its “black-box” nature and limited interpretability remain critical obstacles. This study aimed to develop an uncertainty-guided deep learning segmentation framework and evaluate its accuracy and clinical utility. Methods We retrospectively enrolled 210 patients with spinal tuberculosis from an initial cohort of 300 screened at three centers, and acquired preoperative FS-T2WI scans. Data from Center 1 (n = 160) were used for five-fold cross-validation, while the remaining 50 cases served as an external test set. We built an improved model on nnU-Net by integrating boundary-aware loss with Monte Carlo Dropout, and compared it against U-Net, Attention U-Net, and TransUNet. The uncertainty threshold was determined through internal cross-validation, and clinical validation followed a within-subject crossover design involving nine physicians. Results On the external test set, the improved model achieved a Dice similarity coefficient of 0.858 and an AUC of 0.912, outperforming all comparative models. Uncertainty correlated strongly and positively with pixel-level error rates (Spearman ρ = 0.74). At a threshold of 0.52, sensitivity for identifying unreliable segmentation was 84.3% and negative predictive value was 89.5%. In the physician validation, overlaying uncertainty maps significantly increased trust scores (3.8 ± 0.7 vs. 3.2 ± 0.8; P < 0.001, Cohen’s d = 0.80) and reduced review time (P = 0.018), with the greatest benefit observed among resident physicians. In cases of obvious AI failure, uncertainty alerts accurately flagged erroneous regions in 80.0% of instances. Conclusion The proposed uncertainty-guided framework improved spinal tuberculosis lesion segmentation accuracy. Pixel-level uncertainty maps reliably identified unreliable predictions and enhanced clinician trust, offering a robust AI decision-support tool for precise surgical planning.

Enhancing 3D MRI-Based Necrotic Core Segmentation in Glioblastoma Using Activation Functions in Deep Learning

Precise brain tumour delineation is vital for therapy protocols and tracking. However, standard Rectified Linear Units (ReLU) struggle to capture subtle necrotic-core variations due to zero-gradient behaviour in the negative domain. To address this, we present a controlled benchmark of 12 activation functions within a fixed Residual 3D U-Net using the Brain Tumour Segmentation (BraTS) 2020 dataset. In the single-run benchmark, Swish achieved the best necrotic-core (NCR) Dice (0.676; +2.0% over ReLU, p < 0.01), while TanhExp attained the highest whole-tumour accuracy (0.879). To test the reliability of these single-run results, the four functions central to our claims were retrained across three random seeds. This analysis confirmed a small but consistent NCR advantage for the smooth and adaptive functions—Swish (0.677 ± 0.003) and PReLU (0.678 ± 0.004) over ReLU (0.661 ± 0.012; pooled p < 0.001)—with Swish among the most stable functions in this region. By contrast, the apparent single-run differences in the enhancing tumour, and the underperformance of PReLU, did not generalise across seeds, indicating that activation-function effects in this task are concentrated in the necrotic core and that single-seed comparisons can be misleading. Crucially, Swish achieved these gains with zero additional trainable parameters and only a ~1% training latency penalty on common hardware. Replacing ReLU with Swish offers a cost-effective, architecture-preserving strategy to improve segmentation reliability and boundary delineation. Ultimately, this zero-cost architectural modification is a promising, preliminary step towards more reliable automated tumour delineation, pending prospective validation on multi-institutional data and expert radiological assessment.

Improved Deep Learning Segmentation of Pediatric Diffuse Midline Gliomas After Treatment.

PURPOSE To develop and validate a pediatric diffuse midline glioma (DMG) auto-segmentation tool optimized for longitudinal treatment response assessment across the disease course. MATERIALS AND METHODS In this multi-institutional retrospective study, we included patients aged 1-30 years with DMG from an institutional pediatric cancer center, BraTS-PEDs 2024, and PNOC007, a prospective trial of radiation followed by peptide vaccine plus poly-ICLC, and we trained nnU-Net-based DMGtracker using expert segmentations from 140 institutional pre- and post-treatment studies and all 261 BraTS-PEDs 2024 pre-treatment studies, using four-sequence multiparametric MRI (T1, T1 post-contrast, T2, and FLAIR). We externally validated the model on 88 annotated PNOC007 studies (n = 49 patients) and compared it with the BraTS-PEDs 2024 winning model using median Dice similarity coefficient (DSC) and relative volumetric difference (RVD) for whole-tumor and contrast-enhancing tumor segmentation using the Wilcoxon signed-rank test. RESULTS Training and internal testing used 153 scans (59 post-treatment) from 74 patients. Incorporating post-treatment data improved internal whole-tumor DSC for our trained model (0.94 [IQR 0.82-0.96] vs 0.93 [0.81-0.96]; p<0.001). On external validation, DMGtracker outperformed the BraTS-PEDs 2024 winning model for whole-tumor segmentation, with higher DSC (0.90 [0.72-0.95] vs 0.81 [0.66-0.90]) and lower RVD (9.6% [3.7%-31.6%] vs 16.8% [7.3%-39.2%]). This advantage was greatest in post-treatment scans (n = 50 scans, DSC 0.90 [0.73-0.94] vs 0.80 [0.58-0.88]; RVD 9.7% [3.8%-26.2%] vs 19.8% [12.6%-39.8%]; p<0.001 for both). In post-treatment scans, DMGtracker achieved clinically acceptable whole-tumor segmentation (DSC > 0.80) in 64.0% of cases, compared with 52.0% for the BraTS-PEDs winner. CONCLUSION Training DMG segmentation models with post-treatment scans substantially improves performance in longitudinal clinical trial imaging, enabling more accurate volumetric tracking and response assessment.

A deep learning algorithm for fully automated volumetric measurement of meningioma burden

Abstract Background We sought to develop a deep learning (DL) model to enable fully automated 3D segmentation and volumetric assessment of meningioma burden with a specific emphasis on generalizing to high-grade and posttreatment meningiomas to improve interobserver variability and decrease reader time investment in tumor response assessment. Methods In total, 450 postcontrast T1-weighted brain MRIs from 104 patients with meningiomas were obtained from Massachusetts General Hospital and Dana-Farber Cancer Institute. The cohort was unique among prior DL segmentation models in that it encompassed meningiomas of all grades, postoperative, and postradiated meningiomas. Preprocessed MRIs and manually generated tumor segmentations were used to train a U-Net with a joint Dice-cross entropy loss function. Results When tested on internal data, our model achieved a median Dice of 0.741 and a median 95th percentile Hausdorff Distance (HD95) of 26 mm on a high-grade test set and a median Dice of 0.848 and median HD95 of 1.41 mm on a test set with low-grade tumors. Lesion-wise metrics were equivalent to global metrics for low-grade tumors, which contained only single lesions, but were substantially lower for high-grade tumors, with a median lesion-wise Dice of 0.45 and median lesion-wise HD95 of 130 mm, reflecting greater difficulty delineating individual high-grade lesions. Our model also generalized well to 1000 studies from 944 patients selected from the public BraTS dataset, achieving a median Dice of 0.923 and median HD95 of 2.24 mm. Conclusions The study produced a model that addresses an unmet need for automated volumetric measurements of meningiomas and created a reliable metric for quantifying meningioma burden. In comparison to prior DL approaches, our model achieved competitive performance on external data and improved Dice scores on high-grade and posttreatment meningiomas. The trained model, volumetric evaluation code, and accompanying documentation are available online at https://github.com/mccle/tumor_segmentation.