Deep learning models have shown strong performance in chest
X-ray pneumonia classification, but their robustness to
adversarial perturbations remains a major concern for
clinical deployment. In this paper, we investigate the
adversarial robustness of a Convnet model for chest X-ray
pneumonia classification under multiple attack settings and
evaluate the effectiveness of adversarial training as a
defense mechanism. The model was trained on a chest X-ray
dataset and evaluated on both an internal test set and an
external RSNA dataset to assess cross-dataset
generalization. We considered several adversarial attacks
across multiple perturbation strengths. In addition,
feature squeezing was analyzed as a lightweight defense
strategy. On the internal test set, the standard model
achieved an accuracy of 0.8686 and recall for the pneumonia
class of 0.997, but its performance dropped sharply under
adversarial perturbations, reaching near-zero accuracy for
stronger iterative attacks. In contrast, the adversarially
trained model preserved strong clean performance, achieving
an accuracy of 0.8958, while improving robustness across
all evaluated attacks. For example, under PGD with
perturbation budget 0.003922, the standard model achieved
only 0.0385 accuracy, but the adversarially trained model
achieved accuracy of 0.8285. These results indicate that
adversarial training provides a substantial robustness
without sacrificing baseline predictive performance.
Overall, our findings highlight the importance of
robustness-aware evaluation in medical imaging and support
adversarial training as a practical strategy for improving
the reliability of chest X-ray classification systems.

Analyzing the interface between tooth tissue and
restorative materials in scanning electron microscopy (SEM)
images is important and innovative approach for evaluating
the quality of dental restorations. However, this is
typically done only descriptively, which is inconsistent
between different examiners. This paper presents an
automated pipeline for segmenting the filling region in
annotated SEM micrographs of teeth restored with
glass-ionomer cement. The pipeline uses color markers
placed along the filling boundary before imaging. Since SEM
images are grayscale, these markers are the only colored
elements present. The saturation channel in HSV color space
is extracted and binarized using Otsu’s thresholding to
detect the boundary line. The filling region is then
removed using directional edge scanning and binary masking.
The resulting masks are used for training a deep learning
model based on the U-Net architecture with a ResNet50
encoder. The model was trained and evaluated using 5-fold
cross-validation on a dataset of 36 images, achieving an
average Dice coefficient of 0.87 and an intersection over
union (IoU) score of 0.79. The trained model was then
tested on 25 previously unseen images and evaluated by a
dental specialist. The results show that automated
segmentation of filling regions in SEM images is sufficient
and could be employed in the reduction of time spent for
manual analysis.

Accurate 3D dental landmark detection is a foundational
step in automated orthodontic planning, aligner design, and
occlusal analysis. We present a comparison of three
lightweight point-cloud encoders, hierarchical PointNet,
PointNeXt, and PointViG, equipped with a unified multi-head
architecture for joint full-mouth tooth presence
prediction, fixed-count landmark localization, and
variable-count cusp tip detection. Both arches are merged
into a single full-mouth point cloud that preserves
interarch geometry. The landmark head uses learned
cross-attention queries indexed by tooth slot and landmark
category, decoded via a differentiable soft-argmax.
Variable-count cusp tips are handled by a dedicated heatmap
channel decoded by k-means clustering, where the number of
clusters is provided by an auxiliary count regression head.
Evaluated under 5-fold cross-validation on the 3DTeethLand
benchmark, PointViG achieves a mean radial error of 1.78 mm
and both mean average precision and mean average recall of
0.548, reaching competitive leaderboard performance without
ensembling or per-tooth spatial conditioning. Among the
three backbones, PointViG’s dynamic graph convolution in
feature space proves decisive: its ability to aggregate
information from semantically related but spatially distant
points yields consistent gains across all landmark
categories, particularly on geometrically complex posterior
teeth.

This work presents a comparative evaluation of simple
recurrent neural networks (RNNs) and Long Short-Term Memory
(LSTM) networks for data-driven modeling of dynamic
nonlinear systems in biomedical engineering. The benchmark
system used in this work is a two-compartment
pharmacokinetic system with Michaelis-Menten elimination.
Both architectures are trained using an identical data set
and training procedure and are evaluated on an unseen test
set to ensure a fair comparison. Model performance is
assessed using multiple error-based metrics and the
coefficient of determination to evaluate both prediction
accuracy and dynamic fidelity, alongside a detailed
analysis of computational performance.

The results indicate that the LSTM outperforms the RNN in
terms of predictive accuracy. In contrast, the RNN
consistently achieves lower inference times and reduced
execution-time variability. These findings highlight a
clear trade-off between modeling accuracy and computational
cost.

The interpretation of complex internal medicine reports
remains a significant challenge for patients, frequently
leading to 'semantic reversals' and diminished treatment
adherence. The primary objective of this research is to
bridge the communication gap between technical clinical
documentation and patient health literacy through an
automated simplification pipeline. HealthOCR-RAG was
developed as a modular framework that integrates Optical
Character Recognition (OCR) for digitizing physical reports
with a privacy-first processing layer for mandatory
anonymization. The system utilizes Retrieval-Augmented
Generation (RAG) to ground large language model (LLM)
explanations in verified medical knowledge bases. The
implementation suggests that combining temporary in-memory
processing with schema-controlled generation can reduce the
risks of unsupported medical content and unnecessary
privacy exposure. The principal contribution of this work
is a safety-aware architecture that transforms noisy
scanned documents into structured, patient-centric
explanations without providing diagnostic advice.

This paper presents a multi-domain feature selection
framework for functional Near Infrared Spectroscopy (fNIRS)
signals to identify the most informative features for
distinguishing neutral and incongruent states during the
Stroop task. A total of 280 features were extracted from
both oxygenated hemoglobin (HbO) and deoxygenated
hemoglobin (HbR) signals. The extracted features were
extracted from three domains: time, frequency, and wavelet.
Feature selection was done by using the Fisher score
method. Results indicate that mean amplitude (time domain)
and wavelet coefficients are the most informative features.
The most relevant channels for classification were in
frontal and frontal-central region. Selected features and
channels have potential for application in distinguishing
cognitive states using fNIRS.

Over the past decades, the global incidence of renal cell
carcinoma (RCC) has been steadily increasing, while early
detection remains challenging due to the asymptomatic
nature of the disease and limitations of current diagnostic
methods. Standard imaging techniques are often insufficient
for accurate tissue characterization at the cellular level,
and definitive diagnosis typically relies on invasive
biopsy procedures. This study proposes a novel,
non-destructive diagnostic approach based on opto-magnetic
imaging spectroscopy (OMIS) combined with machine learning
(ML) algorithms for the classification of human kidney
tissue. Unlike the conventional OMIS setup that uses white
light, this work introduces ultraviolet (UV) illumination
to enhance sensitivity to superficial tissue structures and
endogenous fluorophores. A total of 40 ex vivo human kidney
tissue samples (20 healthy and 20 cancerous) were analyzed.
Characteristic OMIS spectra were extracted and used as
input features for multiple ML models, including K-Nearest
Neighbors (KNN), Support Vector Machine (SVM), Random
Forest (RF), and Naive Bayes (NB) classifiers. Model
performance was evaluated using stratified k-fold
cross-validation and standard metrics. The best performance
was achieved using the SVM classifier, with an accuracy of
80% and a Cohen’s kappa of 0.60. Although this performance
does not yet meet clinical standards, the results
demonstrate the feasibility of distinguishing between
healthy and cancerous kidney tissue. The findings suggest
that UV-enhanced OMIS combined with machine learning
represents a promising step toward an objective diagnostic
tool for kidney cancer.

In this paper, a low-coherence interferometric method for
non-contact refractive index measurement is experimentally
investigated with emphasis on biomedical samples. The
experimental setup is based on a fiber-optic configuration,
where the optical path difference (OPD) is obtained from
the recorded interference spectrum using Fast Fourier
Transform (FFT). The method was first tested on several
representative biomedical liquids, including buffer
solutions, artificial sweat, artificial saliva,
phosphate-buffered saline (PBS), and sodium chloride
solutions, to assess the measurement procedure and
performance. The method was then applied to characterize
stereolithography (SLA) photopolymers (NextDent Ortho IBT
and Ortho Flex), which are relevant for biomedical and
microfluidic applications but lack refractive index data in
the near-infrared range. The results demonstrate the
potential application of the proposed method for optical
characterization of biomedical fluids and novel
photopolymers, while also indicating limitations related to
sample geometry and measurement uncertainty.

This work presents a novel 3D-printed stabilization
platform for improving the operational stability and
repeatability of open microfluidic systems. The platform
introduces mechanical confinement and controlled sealing
through an adjustable stabilization platform, enabling
improved control of fluid media while preserving the
accessibility and ease of fabrication of open systems. A
cross-junction droplet generator fabricated via SLA 3D
printing in combination with the fabricated stabilization
platform was used as a proof-of-concept to evaluate system
performance. Experimental results demonstrate stable
droplet generation across a range of flow values, with
droplet size increasing from approximately 100 μm to 450 μm
and generation frequency rising from 1.72 Hz to 17.73 Hz
with the increase in flow rate. Additionally, a hysteresis
effect in droplet formation was observed, most likely due
to residual fluid remaining within the channels. Despite
these effects, the system achieved repeatable operation and
consistent droplet generation under controlled conditions.
The proposed system effectively bridges the gap between
open and closed microfluidics by introducing structural
stabilization without compromising on modularity, ease and
cost of fabrication, offering a practical pathway toward
more reliable open microfluidic platforms for biomedical
applications.