Background/Objectives: The automatic and accurate segmentation of mass lesions in mammograms is a critical step in the computer-aided diagnosis workflow. This study evaluates the mask region-based convolutional neural network (Mask R-CNN), a robust deep learning model for instance-level segmentation, highlighting its growing potential in medical imaging. Methods: Following configuration optimization, the model is applied to mammography images from the CSAW-CC and TOMPEI-CMMDs, which provide masks annotated by breast radiology experts, to be used as a ground truth. To quantify shape and contour discrepancies with respect to the radiologists’ annotations, performance is evaluated with both overlap- and distance-based metrics. Furthermore, the model is tested on synthetic mammograms and benchmarked against classical segmentation methods (Otsu thresholding, gradient vector flow active contour, and Chan–Vese segmentation), Fuzzy c-means clustering, and the state-of-the-art Medical Segment Anything (MedSAM) model. Results: By leveraging transfer learning, a limited dataset suffices for stable training, with the network converging by the 10th epoch. The optimal configuration achieves a bounding box average precision (BBox AP) of 0.80 and a segmentation average precision (Segm AP) of 0.58. Evaluated on a test set of real mammograms, the model correctly identifies the lesions, yielding an average Dice similarity coefficient (DSC) of 0.87 and an average intersection over union (IoU) of 0.77, performance aligned with the current state of the art. Conclusions: Mask R-CNN has emerged as a reliable and robust model for breast lesion segmentation; compared to the other methods evaluated in this study, it demonstrates superior performance across both overlap- and distance-based metrics. The model exhibits high DSC and IoU scores even on synthetic mammograms with a priori known lesion profiles, which can be used as an objective ground truth to mitigate annotation bias from inter-operator variability among radiologists.
Background/Objectives: A common diagnosis for lumbar disc herniation is an image taken with Magnetic Resonance Imaging (MRI). Due to the need for expertise during manual interpretation of MRI scans by specialists, the process is very time consuming. The use of a machine learning (ML)-based pathway to classify lumbar MRI images as either “normal” or “herniated” represents an efficient way to analyse MRI scans seeking similar pathologies. Methods: The proposed methodology transforms original lumbar MRIs (from colour to greyscale), resizes them, filters, and normalises them (using bilateral algorithms) before conducting extraction of features using PCA-based decomposition. A total of ten classical ML classifiers were evaluated using five-fold outer and three-fold inner nested cross-validation, with the number of PCA components selected exclusively within the inner folds. Results: Extra Trees achieved the highest F1-score of 85.86% (95% CI: 84.16–87.44%) and the highest ROC-AUC of 90.96% (95% CI: 89.27–92.54%), with an accuracy of 84.29%, precision of 82.03%, recall of 90.07%, and specificity of 77.79%. Random Forest achieved the highest accuracy of 84.36% (95% CI: 82.40–86.26%), although its overall performance was comparable to that of Extra Trees. Conclusions: The findings support the potential of traditional machine learning classifiers for binary classification within the evaluated image-level dataset, while maintaining a relatively simple design and computational efficiency. No direct empirical comparison with deep learning models was conducted in this study.
Background: Adjuvant radiotherapy (RT) after gross-total resection of World Health Organization (WHO) grade 2 meningioma remains an individualized decision in the absence of definitive randomized evidence. We aimed to identify the clinical, radiological, and pathological factors associated with observed adjuvant RT allocation after Simpson grade I resection and to evaluate the discriminatory performance of proliferative biomarkers in real-world multidisciplinary practice. Methods: This single-center, retrospective cohort included 65 patients, of whom 56 undergoing Simpson grade I resection constituted the primary analytic cohort. Associations with observed RT allocation were evaluated using Firth-penalized logistic regression. Ki-67 and mitotic-index discrimination was summarized using receiver operating characteristic (ROC) analysis, and Ki-67 threshold stability was assessed in 5000 bootstrap resamples. Brain invasion was evaluated in a documented-case sensitivity analysis. Results: Adjuvant RT was administered to 24 of 56 patients (42.9%). In models adjusted for age, sex, neutrophil-to-lymphocyte ratio, and tumor size, each 5-percentage-point increase in Ki-67 was associated with higher odds of RT allocation (OR 7.10, 95% CI 2.20–22.94; p = 0.001), as was each increase of 5 mitoses per 10 high-power fields in mitotic index (OR 5.81, 95% CI 2.01–16.75; p = 0.001). Ki-67 showed greater apparent discrimination than mitotic index (area under the curve [AUC] 0.925 vs. 0.801; DeLong p = 0.008). The apparent Youden-optimal Ki-67 threshold was 14%, but thresholds of 12%, 14%, and 15% were selected in 33.4%, 41.9%, and 23.1% of bootstrap samples. Among 44 patients with documented brain-invasion status, Ki-67 remained associated with RT allocation after adjustment for brain invasion (OR 4.53, 95% CI 1.77–11.63; p = 0.002). Conclusions: Ki-67 and mitotic index were associated with observed RT allocation, but both markers were already available to the tumor board. These findings describe institutional treatment-allocation patterns and do not establish RT benefit or validate a Ki-67 threshold for recommending treatment.
Background/Objectives: Eosinophilic esophagitis (EoE) is a chronic disease of the esophagus caused by an abnormal type 2 (Th2) immune response. It is a leading cause of dysphagia in both pediatric and adult populations. Endoscopy with biopsy remains the standard approach for diagnosis, but its invasive nature and need for repeated procedures have prompted investigation of alternative and complementary approaches. Methods: A narrative review was conducted to synthesize evidence on established and emerging modalities for EoE diagnosis and disease assessment, including minimally invasive sampling techniques, functional and structural assessment, mucosal impedance, molecular diagnostics, tissue-based assessment, and artificial intelligence (AI)-assisted approaches. Results: Minimally invasive approaches may facilitate diagnosis or longitudinal monitoring, while complementary modalities provide assessment of structural, functional, histologic, and molecular features of disease. In retrospective image datasets, AI models have demonstrated high performance for automated eosinophil quantification and detection of endoscopic features, although none has been prospectively validated in routine clinical practice. Conclusions: Despite these advances, widespread adoption remains limited by cost, accessibility, and variability in validation across diverse populations. Further prospective and external validation is needed to establish the role of these modalities in mainstream practice.
Background: Total hip arthroplasty (THA) and the diagnosis of periprosthetic joint infection (PJI) have each acquired a technology that measures with high precision: intraoperative navigation, which reports component position to within a few degrees, and molecular methods, which identify the infecting organism and the host response at genomic resolution. Objective: To establish whether the measured accuracy has been shown to reach the patient as a clinical outcome. Methods: This is a narrative review, reported according to SANRA. MEDLINE, Scopus, and Web of Science databases were searched from inception to 31 August 2026. English-language studies in humans were eligible if they reported positional or diagnostic accuracy, a patient-level clinical or economic outcome after THA or after treatment of PJI or osteomyelitis, or a factor modifying the relation between the two. Animal, in vitro, and cadaveric studies, conference abstracts, and reports without extractable data were excluded. Forty studies were included. No pooled estimate was calculated, and no new statistical analysis was performed. Results: Navigation reduced positional outliers consistently, but in two matched claims-based cohorts of patients with morbid obesity (4499 and 5835 patients), it was not associated with lower rates of PJI, dislocation, periprosthetic fracture, mechanical loosening, or revision within two years. Candidate immunogenetic variants in osteomyelitis have never been tested against a treatment decision. Metagenomic sequencing was no more accurate than culture (sensitivity 63.2% versus 68.4%; specificity 80.0% versus 95.0%) except after antibiotic exposure. Reported infection rates and test accuracy both change with the case definition applied. Conclusions: In both settings, the limiting factor is the design of the studies linking measurement to outcome rather than the precision of the measurement.
