Anitha Hoblidar

Goal: Objective: Renal Cell Carcinoma (RCC) is a heterogeneous cancer with multiple histological subtypes that require individualized clinical management. Once implemented, timely and appropriate subtype delineation from imaging data can lead to improved diagnostic accuracy and intervention decisions. Methods: Study compares 3 convolutional neural networks (CNN) architectures, AlexNet, GoogleNet, VGG16, Resnet18 and Densenet to determine which approach has the best performance for RCC subtype classification of Contrast-Enhanced computed tomography (CECT) images. The original dataset of 590 images from 90 patients was augmented using anatomically constrained augmentation (rotation ±5° and 2D scaling) to generate a total of 3469 unique images, with allocation of training, validation, and testing sets was 70:15:15, respectively. Each architecture was fine tuned to employ an optimized learning rate, dropout regularization, and L2 penalties to minimize overfitting. Results: The quantitative metrics for each architecture, displayed in Table IV, show that VGG16 achieved the maximum level of 99.21% accuracy but demonstrated some levels of overfitting behavior at the lower learning rates. GoogleNet achieved commendable accuracy levels of 97.67% accuracy after epoch optimization, but performance levels were highly sensitive across the learning rate. AlexNet, in comparison demonstrated the most stable convergence rates and achieved strong levels of generalization within all three large learning rates, achieving levels of 98.84% accuracy, and processing times were far lower than for either of the other networks. Conclusion: this work demonstrates that GoogleNet is a reliable and cost efficient implementation of a CNN for a clinical setting that enables rapid accurate clinical decision making from imaging data for the specific subtype of RCC. Significance: These findings reinforce the potential of tailored deep learning frameworks for early RCC subtype classification and encourage further validation on larger, multi-institutional datasets to enhance clinical applicability.

The electrical activities in living tissues are caused by ionic movements along cell membranes, particularly in excitable tissues like neurons, cardiac cells, and skeletal muscles. The heart muscles produce currents that flow through the tissue in the body’s volume conductor. These currents generate potentials and magnetic field that are measurable at and beyond the thorax surface using detectors. Researchers have been studying these electrical activities by modeling their electrophysiology. This research work discusses the different approaches to model electrical activities of the heart. One challenge in this area is developing algorithms to visualize or localize cardiac anomalies non-invasively. This requires a prior forward model that represents the spatial relationship between the heart and the detectors. Traditional forward models assume the locations and strengths of the current sources but not their orientations, leading to potential inaccuracies or misinterpretations in inverse calculations. Further, this work introduces a novel method for constructing the forward matrix by constraining prior heart vectors scaled with time varying information of activation sequences. The resulting spatial matrix is used to locate hidden abnormal sources in the myocardium through solving the inverse problem. The inverse problem is extended to identify ventricular anomalies by modeling ruptured nodes in specific heart regions.

Background: Magnetocardiography (MCG) is a non-invasive technique that measures magnetic fields generated by cardiac electrical activity, offering superior spatial resolution compared to traditional electrocardiography (ECG) for localizing arrhythmic sources. Accurate dipole source localization remains a critical challenge in cardiac electrophysiology. Methods: This work developed a comprehensive computational framework integrating the Boundary Element Method (BEM) for forward electromagnetic modeling with the Nelder-Mead simplex algorithm for inverse source optimization. A realistic cardiac geometry was formed by extracting a standard dataset from ECGSIM software and strategically selected 16 nodes. An experimentally acquired 6 Hz pulse signal was used as the temporal pattern for current dipole sources; appropriate amplitude scaling was applied in the implementation to achieve physiologically realistic magnetic field strengths of 10–100 pT consistent with clinical MCG measurements. Magnetic field distributions were computed at 25 sensor grid locations configured in a planar array. Forward field calculations are performed using the Biot-Savart law, while the inverse problem minimized the squared error between measured and simulated magnetic fields by iterative optimization of the dipole position. Results: Proposed model achieves sub-millimeter localization precision across multiple validation scenarios. Comprehensive testing in 16 locations yields a mean localization error of 0.715 ± 0.192 mm (range: 0.318–0.968 mm), which with algorithmic convergence was achieved in 15-46 iterations. Under noise-free conditions, optimization framework reliably converged to solutions meeting the clinically relevant 1 mm accuracy criterion. Noise analysis using signal-to-noise ratios from -10 to 20 dB shown clinically acceptable performance (< 5 mm error) at SNR ≥ 5dB, with mean errors of 0.949 mm at 20 dB and 3.025 mm at 10 dB representing typical shielded and unshielded clinical environments. Comparative analysis against Levenberg-Marquardt, Genetic Algorithm, and Particle Swarm Optimization methods confirmed Nelder-Mead’s superior combination of localization accuracy (0.715 mm vs. 2.984-76.517 mm for other existing methods) and 100% convergence reliability across all test configurations. Conclusions: This BEM-Nelder-Mead framework shows reliable cardiac dipole localization with acceptable accuracy. MATLAB implementation offers reproducibility and adaptability for various cardiac source configurations, demonstrating potential for non-invasive arrhythmia diagnosis and therapeutic guidance.

Purpose: Accurate pedicle screw placement is crucial for ensuring safety during spine surgery. This study aims to develop and validate a feature-based 3D to 2D registration framework to integrate pre-operative MR with intra-operative X-ray images for estimating the pedicle screw trajectory and achieving robust 3D visualization. The method aims to overcome intensity mismatches and tissue non-correspondence issues inherent to multimodal registration. Methods: A feature-based registration framework was proposed, leveraging vertebral endplates as key features due to their consistency after pedicle screw insertion. The segmented MR images were forward-projected and registered to intra-operative X-ray images using a binary image matching similarity measure, with optimization achieved through CMA-ES. Validation was performed by mapping the registered trajectory position onto pre-operative MR images and further evaluating the accuracy through 3D to 3D registration of pre-operative MR and post-operative CT images. Results: The framework effectively estimated the pedicle screw trajectory by registering intra-operative images onto pre-operative MR, with errors consistently below 2.5 mm for entry point, endpoint, and directional trajectory. The 3D to 3D registration validated the accuracy, demonstrating precise alignment between trajectory positions across modalities. Conclusion: The proposed feature-based registration framework provides fast and accurate 3D trajectory information. This method improves surgical precision and offers robust evaluation metrics, making it a reliable tool for spine navigation and clinical diagnosis during pedicle screw placement procedures.

Recent advances in non-invasive cardiac imaging techniques have highlighted the potential of magnetocardiography (MCG) as a powerful diagnostic tool. However, existing approaches face significant challenges in accurately localizing and characterizing cardiac electrical sources. This study presents an innovative methodology for cardiac source localization using magnetocardiography based on the magnetoimpedance principle, addressing these limitations through an experimental framework and sophisticated computational approaches. This work has established a novel lead field matrix computation methodology that shows the relationship between source currents and resulting magnetic fields. The forward problem was solved using known source configurations and validated against experimental measurements. For the inverse problem, three mathematical approaches were implemented: Minimum Norm Estimation (MNE), empirical covariance, and Gaussian Process Variance methods. Results demonstrate that the overall experimental setup successfully replicates key aspects of cardiac electromagnetic activity, with the MI sensor array providing sufficient spatial and temporal resolution for accurate source localization. The novel lead field computation method significantly improved source reconstruction accuracy compared to conventional approaches. This work establishes a foundation for cost-effective high-sensitivity magnetocardiographic systems that could enhance clinical cardiac diagnostics without the need for magnetically shielded environments.

The human heart produces electrical signals to contract and relax its muscles that helps in its blood pumping activities. These electrical impulses give rise to electric potentials on the body surface and tiny magnetic field around the thorax. These functional activities can be investigated using Electro/Magnetocardiogram (E/MCG). The more challenging task in the E/MCG research is to image the cardiac dysfunctions in three dimensions not at the surface level but at the source level and this is called the inverse problem. To solve this, one has to model a generic structure of the discretised heart enclosed in the thorax mesh and their spatial relation with location of the MCG detectors, called forward problem. In this current research, sources in a homogeneous volume conductor model is used in the construction of forward problem. A novel algorithm is implemented that uses Vectocardiography (VCG) signals in the forward problem of MCG. Another objective of this paper includes the utilization of dynamic lead field based on VCG orientations in the inverse problem. In this study, the ill-posed problems are solved using Bayesian approach and the results are compared with the deterministic approach for measurements on noise signals. The analysis revealed that the proposed algorithms with VCG priors (that are extracted from the VCG signals) in the Bayesian framework significantly improved MCG source localization in cases of Myocardial Ischemia. Analysis from the study showed that the proposed algorithms with VCG priors in probabilistic methods significantly captured a good region of spread of the reconstructed borders of inverse solutions. The average spread for deterministic methods was around 3.5 - 4.3cm in the diseased cases with simulated true ruptured region of spread being 2.5cm. In contrast, Bayesian methods and total variation methods with VCG priors reduced the spread to 2.9 to 3.03cm, respectively. The introduction of VCG signals in the forward problem of MCG not only increases the accuracy of cardiomagnetic imaging but also provides a path for more reliable diagnostic tools in cardiology.

Goal: This study aims to explore the temporal dynamics of functional connectivity in drug-resistant focal epilepsy, focusing on Temporal Lobe Epilepsy (TLE) and Extra-Temporal Lobe Epilepsy (ETLE), using magnetoencephalography (MEG). Methods: Temporal metrics such as Change Between States, Entropy of Transition Patterns, Entropy of Transition Probabilities, Dwell Time, Stability, and Max L1 Distance derived from dynamic functional connectivity matrices were analyzed across eight frequency bands (delta, theta, alpha, beta, low gamma, mid gamma, high gamma and broadband) in TLE and ETLE patients. Results: Significant differences were observed between TLE and ETLE. ETLE exhibited more widespread and unpredictable connectivity transitions, while TLE demonstrated localized and structured patterns. Entropy metrics indicated higher randomness in ETLE, and dwell time analysis revealed shorter state persistence in ETLE compared to TLE. Conclusions: The findings highlight the potential of MEG-based temporal connectivity metrics in characterizing network disruptions in focal epilepsy.

Accurate identification of fiducial markers, such as the left pre-auricle (LPA), right pre-auricle (RPA), and nasion, is critical for head localization, source reconstruction, anatomical alignment, and ensuring reproducibility in functional and structural neuroimaging studies. However, manual annotation of these landmarks is time-consuming, prone to human error, and further complicated by variations in MRI acquisition protocols and incomplete head coverage. In this work, we propose a deep learning-based framework using autoencoders to automatically detect fiducial markers in 3D brain MRI scans. The method integrates a structured preprocessing pipeline, including intensity normalization, spatial resampling, and noise reduction, to ensure data consistency. A tailored autoencoder architecture is then trained to reconstruct MRI slices while learning the spatial characteristics of MRI slices at fiducial points across axial, coronal, and sagittal planes. Fiducial localization is determined based on the slice with the lowest reconstruction error, ensuring precise identification. Our approach significantly enhances registration accuracy, achieving an average ground truth Euclidean distance of 1.42 mm for LPA and 1.89 mm for RPA, with low variance across a dataset of 500 MRI volumes. The results demonstrate that automated fiducial detection can reduce manual intervention, improve anatomical alignment, and enhance the robustness of neuroimaging workflows. This method offers a scalable and efficient solution for clinical and research applications, facilitating more reliable head modeling, source localization, and multimodal image integration.