Flakë Bajraktari

Abstract

This study explores the potential of using a sensor system and deep learning algorithm for automatic recognition and classification of human movement intentions to enhance the control of active orthopedic aids such as orthoses, prostheses, and exoskeletons. The sensor system consists of four inertial measurement units (IMU) attached to the thighs, lower back, and chest of a person, measuring acceleration and rotational velocity in three axes each. A dataset was generated from 20 healthy individuals performing various movements from a standing position, and preprocessed and segmented into time windows. The data was fed into a convolutional neural network (CNN) capable of classifying the time windows into seven classes. The network achieved a classification accuracy of up to 82 %. The results demonstrate the potential of the proposed system for successful application in the control of assistance systems, which can improve the operability and effectiveness of orthopedic aids.

Abstract

Automatic recognition of surgical workflows is a complex yet essential task of context-aware systems in the operating room. However, achieving high accuracy in phase recognition remains a challenge due to the complexity of surgical procedures. While recent deep learning models have made significant progress, individual models often exhibit limitations—some may excel at capturing spatial features, while others are better at modeling temporal dependencies or handling class imbalance. This study investigates the use of ensemble learning to combine the complementary strengths of diverse architectures, aiming to mitigate individual model weaknesses and improve performance in surgical phase recognition using the Cholec80 dataset. A variety of advanced deep learning architectures was integrated into a single ensemble. Models were carefully selected and tuned to ensure diversity, resulting in a final set of 15 unique ensembles. Ensemble strategies were explored to determine the most effective method for combining the distinct models. The results demonstrated that ensemble learning significantly improved performance. Among the ensemble strategies tested, majority voting achieved the highest F1-score, followed by the proposed artificial neural network StackingNet. Ensembles with high model diversity showed superior performance compared to those with lower diversity. The optimal ensemble configuration integrated top performing models from different architectures, leading to improvements in accuracy, F1-score, and Jaccard Index by 1.48%, 3.68%, and 5.43%, respectively, compared to the best individual models. This study demonstrates that ensemble learning can substantially enhance surgical phase recognition by leveraging the complementary strengths of diverse deep learning models. Ensemble size, diversity, and meta-model selection were identified as key factors influencing performance. The resulting improvements translate into clinically meaningful benefits by enabling more reliable context-aware guidance, reducing misclassifications during critical phases, and improving surgeons’ trust in artificial intelligence (AI) systems.

Abstract

In this paper, the development of a cost-effective assistance system for venipuncture is presented. The system locates forearm veins through near-infrared imaging, depth estimation, deep learning segmentation, and 3D reconstruction. A single-board computer was integrated with two infrared cameras and two 760 nm near-infrared (NIR) LEDs to capture and process stereo images. The depth estimation was achieved through stereo triangulation. A deep learning model based on the U-Net architecture with an attention mechanism and a training dataset of 900 images from 40 participants was used for vein segmentation. Depth information and segmented veins were combined to enable a 3D visualization of the veins. The results show a Jaccard-Score of 92.80 % for vein segmentation and an average reprojection error of 0.48 pixels for the 3D reconstruction.

Abstract

Despite being among the most common medical procedures, needle insertions suffer from a high error rate. Impedance measurements using electrode-equipped needles offer promise for improved tissue targeting and reduced errors. Impedance visualization usually requires an extensive pre-measured impedance dataset for tissue differentiation and knowledge of the electric fields contributing to the resulting impedances. This work presents two finite element simulation approaches for both problems. The first approach describes the generation of a multitude of impedances with Monte Carlo simulations for both, homogeneous and inhomogeneous tissue to circumvent the need to rely on previously measured data. These datasets could be used for tissue discrimination. The second method describes the simulation of the spatial sensitivity distribution of an electrode layout. Two singularity analysis methods were employed to determine the bulk of the sensitivity within a finite volume, which in turn enables consistent 3D visualization. The modeled electrode layout consists of 12 electrodes radially placed around a hypodermic needle. Electrical excitation was simulated using two neighboring electrodes for current carriage and voltage pickup, which resulted in 12 distinct bipolar excitation states. Both, the impedance simulations and the respective singularity analysis methods were compared with each other. The results show that the statistical spread of impedances is highly dependent on the tissue type and its inhomogeneities. The bounded bulk of sensitivities of both methods are of similar extent and symmetry. Future models should incorporate more detailed tissue properties such as anisotropy or changing material properties due to tissue deformation to gain more accurate predictions.