Pablo Menéndez Fernández-Miranda

The reliable deployment of artificial intelligence systems in medical imaging requires high diagnostic performance, robustness and interpretability. In this study, we developed and evaluated two automated frameworks for binary classification of shoulder radiographs (XRs) using deep learning (DL) and hybrid DL–machine learning (ML) approaches. A convolutional neural network (CNN) based on a fine-tuned VGG19 architecture was trained end-to-end on a large, balanced dataset of 4,268 shoulder XRs. In parallel, hybrid models were constructed by extracting deep feature representations from the trained network and combining them with traditional ML classifiers. Model performance was evaluated on independent internal ( n = 480) and external ( n = 308) validation sets. Both approaches achieved high discriminative performance. Paired comparison of Receiver Operating Characteristic (ROC) curves using the DeLong test revealed no statistically significant differences between the end-to-end CNN and the hybrid CNN–ML pipeline for either internal validation (AUC 0.956 vs. 0.961) or external generalization (AUC 0.940 vs. 0.942). Model interpretability was assessed using Grad-CAM and SHAP values. Our results suggest that while both frameworks are robust, the end-to-end DL approach offers a more streamlined workflow and more direct visual explainability via saliency maps. These findings support the potential of AI-based tools for shoulder XR analysis; however, prospective real-world validation, assessment under routine prevalence conditions, and direct comparison with human readers are still needed before clinical integration can be established.

Otosclerosis is a disease affecting the middle and inner ear, characterized by abnormal bone remodeling that leads to stapes fixation and progressive hearing loss. Although high-resolution computed tomography (HRCT) is the standard imaging modality for diagnosis, its sensitivity is limited, with a high false-negative rate (FNR). This study investigates the use of radiomics and machine learning (ML) to improve diagnostic accuracy. HRCT scans from 99 subjects (48 otosclerosis, 51 controls) were analyzed, focusing on the stapes, antefenestral region (AF), and oval window (OW). From each scan, 6048 radiomic features were extracted and reduced to 1317 through feature selection. Statistical analyses and ML modeling were performed using the selected features. Sixty-seven biomarkers showed significant differences between cases and controls, primarily in the AF (56) and stapes (11); none were found in the OW. Both the AF and stapes exhibited increased heterogeneity in otosclerosis, reflecting the bone remodeling process. A reduction in the stapes' major axis was also observed, possibly related to torsional deformation. Image transformation filters enhanced disease visibility. Among several ML classifiers tested, L2-regularized logistic regression performed best, achieving an AUC of 0.90 ± 0.06, thereby enhancing the diagnostic accuracy reported in some studies for radiologists. Hierarchical clustering of the most predictive features further confirmed their strong discriminative power. Our findings highlight the potential of radiomics and ML to standardize otosclerosis diagnosis, reduce FNR, and support surgical decision-making. Future studies should validate these results using larger cohorts and advanced imaging technologies such as Photon-Counting CT.

This study aimed to evaluate whether voxel spacing-aware radiomic feature extraction improves reproducibility, variability, and discriminative performance compared to conventional preprocessing methods in anisotropic CT data. A curated cohort of 685 pulmonary nodules from the LIDC-IDRI dataset was analyzed. Three preprocessing strategies-no resampling, isotropic resampling, and voxel spacing-aware extraction-and one postprocessing approach, voxel spacing weighting, were systematically compared. A modified version of PyRadiomics was developed to compute texture and shape features directly from native images while incorporating physical voxel dimensions without interpolation. Among the 94 extracted features, spacingaware preprocessing improved reproducibility in 58 features, reduced variability in 48, and enhanced univariate discrimination in 37. An ensemble feature selection combining six statistical and machine learning methods identified between 18 and 20 robust features per method. Logistic Regression models trained with spacing-aware features achieved the highest composite performance score (1.498), balancing discrimination and generalizability. SHAP interpretability analysis confirmed the clinical relevance of selected geometric and texture features. Overall, voxel spacing-aware preprocessing preserved native spatial structure and mitigated the effects of voxel geometry and interpolation artifacts, yielding stable and clinically interpretable radiomic features. These findings support the adoption of spacing-aware pipelines in heterogeneous and multi-center CT radiomics studies to enhance feature robustness and reproducibility.

A data-centric approach prioritizing data quality enabled the development of a generalizable deep learning system that accurately detected and segmented brain metastases of all sizes on MRI scans.

Generalization of deep learning (DL) algorithms is critical for the secure implementation of computer-aided diagnosis systems in clinical practice. However, broad generalization remains to be a challenge in machine learning. This research aims to identify and study potential factors that can affect the internal validation and generalization of DL networks, namely the institution where the images come from, the image processing applied by the X-ray device, and the type of response function of the X-ray device. For these purposes, a pre-trained convolutional neural network (CNN) (VGG16) was trained three times for classifying COVID-19 and control chest radiographs with the same hyperparameters, but using different combinations of data acquired in two institutions by three different X-ray device manufacturers. Regarding internal validation, the addition of images from an external institution to the training set did not modify the algorithm’s internal performance, however, the inclusion of images acquired by a device from a different manufacturer decreased the performance up to 8% (p < 0.05). In contrast, generalization across institutions and X-ray devices with the same type of response function was achieved. Nonetheless, generalization was not observed across devices with different types of response function. This factor was the key impediment to achieving broad generalization in our research, followed by the device’s image-processing and the inter-institutional differences, which both reduced generalization performance to 18.9% (p < 0.05), and 9.8% (p < 0.05), respectively. Finally, clustering analysis with features extracted by the CNN was performed, revealing a substantial dependence of feature values extracted by the pre-trained CNN on the X-ray device which acquired the images.

Nystagmus induced by applying an intense vibratory stimulus to the skull (SVIN) indicates vestibular functional asymmetry. In unilateral vestibular loss, a 100 Hz bone-conducted vibration given to either mastoid immediately causes a primarily horizontal nystagmus. The test is performed in darkness to avoid visual fixation (VF) but there are no data about how much VF affects the often-intense SVIN. The aim is to analyze the amount of reduction in SVIN when VF is allowed during testing. Thus, all patients seen in a tertiary hospital for vertigo or dizziness with positive SVIN were included. SVIN was recorded for 10 s for each condition: without VF (aSVINwo) and with VF (aSVINw). We obtained an aSVINwo and an aSVINw as average slow-phase velocities (SPV) without and with VF. VF index (FI_SVIN) was calculated as the ratio of SPV. Among the 124 patients included, spontaneous nystagmus (SN) was found in 25% and the median slow phase velocity (mSPV) (without VF) of SN was 2.6 ± 2.4°/s. Mean FI_SVIN was 0.27 ± 0.29. FI_SVIN was 0 in 42 patients, and FI_SVIN between 0 and 1 was found in 82 (mean FI_SVIN 0.39 ± 0.02). Fixation suppression was found in all patients with SVIN in cases of peripheral vestibulopathy. FI_SVIN clearly delineates two populations of patients: with or without a complete visual reduction in nystagmus.

Nystagmus induced by applying an intense vibratory stimulus to the skull (SVIN) indicates vestibular functional asymmetry. In unilateral vestibular loss, a 100-Hz bone-conducted vibration given to either mastoid immediately causes a primarily horizontal nystagmus. Test is performed in darkness to avoid visual fixation (VF) but there is no data about how much VF affects the often-intense SVIN. The aim is to analyze the amount of reduction of SVIN when VF is allowed during testing. Thus, all patients seen in a tertiary hospital for vertigo or dizziness with positive SVIN were included. SVIN was recorded for 10 seconds for each condition: without VF (aSVINwo) and VF (aSVINw). We obtained an aSVINwo and aSVINw as average slow-phase velocities (SPV) without and with VF. VF index (FISVIN) was calculated as the ratio of SPV. Among the 124 patients included, spontaneous nystagmus (SN) was found in 25% and the mSPV (without VF) of SN was 2.62.4º/s. Mean FISVIN was 0.270.29. FISVIN was 0 in 42 patients and >0 FISVIN <1 in 82 (mean FISVIN 0.390.02). Fixation suppression was found in all patients with SVIN in cases of peripheral vestibulopathy. FISVIN clearly delineates two populations of patients: with or without complete visual reduction of nystagmus.

Deep learning is nowadays at the forefront of artificial intelligence. More precisely, the use of convolutional neural networks has drastically improved the learning capabilities of computer vision applications, being able to directly consider raw data without any prior feature extraction. Advanced methods in the machine learning field, such as adaptive momentum algorithms or dropout regularization, have dramatically improved the convolutional neural networks predicting ability, outperforming that of conventional fully connected neural networks. This work summarizes, in an intended didactic way, the main aspects of these cutting-edge techniques from a medical imaging perspective.