Prathiksha Padmanabha

Ms. Prathiksha Padmanabha

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
27 November 2025 in Healthcare
Verified
Journal Article
Open access
PDF file
Kasunika GurugeKasunika Guruge
Kasunika Guruge
PPPrathiksha Padmanabha
Prathiksha Padmanabha
HHH. M. K. K. M. B. Herath
H. M. K. K. M. B. Herath
NVNuwan Madusanka Vithanage
Nuwan Madusanka Vithanage
HPHi-Joon Park
Hi-Joon Park
BLByeong-il Lee
Byeong-il Lee
Background: Precise localization of acupuncture points (acupoints) is crucial for the clinical success of Traditional Korean Medicine (TKM). Traditional methods that rely on visual inspection and palpation are subjective and prone to inter- and intra-observer differences, making standardization challenging. The progress of data-driven localization techniques is also limited by the scarcity of annotated datasets and inconsistent labeling quality. Objective: This study presents MetaAcuPoint, a synthetic dataset created to overcome these limitations by providing high-fidelity, anatomically consistent hand images for acupoint localization. Methods: MetaAcuPoint was generated using MetaHuman avatars within Unreal Engine, resulting in 900 RGB hand images. Anatomically aligned, bone-attached sockets were implemented for five diagnostically relevant hand acupoints, ensuring millimeter-level precision and spatial consistency across various hand poses. Dataset validity was assessed by training a high-resolution network (HRNet-W48) within the MMPose framework and testing its performance on real-world forearm images. Results: The synthetic-trained model achieved a mean distance error (MDE) of 5.67 ± 3.13 pixels, closely aligning with the real-data baseline at 4.81 ± 2.85 pixels. Adding synthetic samples to real data further enhanced performance (MDE: 4.95 pixels). In contrast, manually annotated synthetic images yielded poorer results (MDE: 12.76 pixels), emphasizing the advantages of automated anatomical annotation. Generalization tests across four external datasets confirmed that the synthetic data-trained model outperformed the real-data-trained model, maintaining higher accuracy (MDE: 5.84–6.45 mm vs. 10.63–15.80 mm). Conclusions: MetaAcuPoint demonstrates the first example of synthetic-to-real generalization for hand acupoint localization. By combining photorealistic rendering with anatomically grounded annotation, the dataset offers a reliable resource to promote standardized, data-driven approaches in acupuncture research and practice.

19 April 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
PPPrathiksha Padmanabha
Prathiksha Padmanabha
HHH. M. K. K. M. B. Herath
H. M. K. K. M. B. Herath
NVNuwan Madusanka Vithanage
Nuwan Madusanka Vithanage
HPHi-Joon Park
Hi-Joon Park
CNChang-Su Na
Chang-Su Na
BLByeong-il Lee
Byeong-il Lee
Accurate anatomical landmark localization in clinical images requires millimeter-level spatial precision, yet whether increasing model scale improves such precision in structured medical imaging tasks remains unclear. Five YOLO26 pose-estimation variants (N, S, M, L, and X) were evaluated on 3679 RGB distal-arm images from 262 participants under a standardized overhead imaging protocol, with five anatomical landmarks annotated across the proximal forearm, mid-forearm, and hand. Localization error was quantified in millimeters using ArUco-marker-based pixel-to-millimeter calibration; all models were initialized from COCO-pretrained weights, fine-tuned under identical conditions, and assessed using COCO-style detection metrics and physically grounded localization error. Detection performance saturated across all scales (mAP@0.5 = 99.5%), while localization performance differed substantially; YOLO26N achieved the lowest mean error (2.76 ± 0.96 mm) and the highest proportion of predictions within 4 mm (88.0%), whereas YOLO26X produced the highest mean error (4.08 ± 2.59 mm) despite a 26.9× higher computational cost. Landmark-wise analysis revealed a consistent proximal-to-distal error gradient, with the largest degradation at anatomically ambiguous proximal landmarks in larger models. These findings suggest that increasing model capacity does not improve clinically meaningful localization precision in structured distal-arm imaging, and lightweight models may offer the most favorable accuracy-efficiency trade-off in resource-constrained clinical settings.

1 September 2025 in Bioengineering
Verified
Journal Article
Open access
PDF file
NVNuwan Madusanka Vithanage
Nuwan Madusanka Vithanage
PPPrathiksha Padmanabha
Prathiksha Padmanabha
Kasunika GurugeKasunika Guruge
Kasunika Guruge
BLByeong-il Lee
Byeong-il Lee
Staining variability in histopathological images compromises automated diagnostic systems by affecting the reliability of computational pathology algorithms. Existing normalization methods prioritize color consistency but often sacrifice critical morphological details essential for accurate diagnosis. This work proposes a novel deep learning framework, integrating enhanced residual learning with multi-scale attention mechanisms for structure-preserving stain normalization. The approach decomposes the transformation process into base reconstruction and residual refinement components, incorporating attention-guided skip connections and progressive curriculum learning. The method was evaluated on the MITOS-ATYPIA-14 dataset containing 1420 paired H&E-stained breast cancer images from two scanners. The framework achieved exceptional performance with a structural similarity index (SSIM) of 0.9663 ± 0.0076, representing 4.6% improvement over the best baseline (StainGAN). Peak signal-to-noise ratio (PSNR) reached 24.50 ± 1.57 dB, surpassing all comparison methods. An edge preservation loss of 0.0465 ± 0.0088 demonstrated a 35.6% error reduction compared to the next best method. Color transfer fidelity reached 0.8680 ± 0.0542 while maintaining superior perceptual quality (FID: 32.12, IS: 2.72 ± 0.18). The attention-guided residual learning framework successfully maintains structural integrity during stain normalization, with superior performance across diverse tissue types, making it suitable for clinical deployment in multi-institutional digital pathology workflows.

19 October 2025
Verified
Book Chapter
Open access
PDF file
PPPrathiksha Padmanabha
Prathiksha Padmanabha
HHH. M. K. K. M. B. Herath
H. M. K. K. M. B. Herath
HPHi-Joon Park
Hi-Joon Park
CNChang-Su Na
Chang-Su Na
MYMyunggi Yi
Myunggi Yi
BLByeong-il Lee
Byeong-il Lee
Accurate localization of anatomical keypoints is fundamental to a variety of clinical and research applications, including acupoint therapy, physical rehabilitation, and biomechanical motion analysis. In this study, we investigate the impact of different deep learning backbone architectures and input image resolutions on the localization accuracy of five clinically important forearm acupoints (LI11, LI10, TE5, LI4, and TE3) using a top-down pose estimation framework. Six backbone models, High-Resolution Network (HRNet), Lite High-Resolution Network (Lite-HRNet), Residual Network (ResNet), Swin Transformer, Visual Geometry Group (VGG) network, and Vision Transformer (ViT), are benchmarked across three input resolutions: 256 × 256, 384 × 384, and 512 × 512 pixels. Localization performance is evaluated using normalized Euclidean distance and its standard deviation, providing a resolution-independent measure of both accuracy and consistency. Among the tested models, HRNet consistently yields the most precise and stable predictions, achieving the lowest average error of 0.009 at the 512 × 512 resolution. Lite-HRNet and ResNet also demonstrate strong performance, particularly at higher resolutions, offering a compelling balance between accuracy and computational efficiency. While ViT exhibits higher error rates at lower resolutions, it significantly improves at 512 × 512, closing the performance gap with CNN-based models. Swin Transformer, however, records the highest average error across all resolutions, indicating its limited suitability for fine-grained localization tasks without additional architectural adaptation. Overall, our findings highlight the importance of backbone selection and resolution scaling in medical keypoint detection tasks. HRNet delivers over 50% lower normalized error compared to the weakest-performing model. These insights provide practical guidance for developing robust, high-precision AI systems for clinical and digital health applications, where anatomical accuracy is essential.

23 August 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
HHH. M. K. K. M. B. Herath
H. M. K. K. M. B. Herath
PPPrathiksha Padmanabha
Prathiksha Padmanabha
NVNuwan Madusanka Vithanage
Nuwan Madusanka Vithanage
MKMailan Arachchige Don Rajitha Kawshalya
Mailan Arachchige Don Rajitha Kawshalya
BLByeong-il Lee
Byeong-il Lee
Parkinson’s disease (PD) is associated with motor impairment and altered cortical dynamics. Although resting-state electroencephalography (EEG) has been widely studied for PD classification, walking EEG remains comparatively underexplored despite its relevance to gait dysfunction. This study developed an EEG-only, leakage-free framework for participant-level classification of PD and healthy controls (HCs) using resting- and walking-state EEG data from the OpenNeuro ds007526 dataset. Following selection, preprocessing, and quality control, 132 participants were retained, including 109 individuals with PD and 23 HCs. EEG recordings were resampled to 128 Hz, filtered between 0.5 and 40 Hz, screened for artifacts, harmonized across conditions, and segmented into overlapping 6 s windows. Classical machine-learning (ML) models used participant-level engineered features, whereas deep-learning (DL) models classified pooled resting-state and walking windows. Extra Trees achieved the highest balanced accuracy (0.71) among the ML, while ShallowConvNet-Lite was the strongest standard DL baseline (0.82). The proposed NeuroAtten-PD model achieved an accuracy of 0.86, a balanced accuracy of 0.84, an F1-score of 0.91, an ROC-AUC of 0.90, and a sensitivity of 0.88. After Holm correction, participant-level paired comparisons indicated significant differences from DeepConvNet-Lite and EEGNet-Lite. These findings demonstrate the feasibility of participant-level PD classification using a pooled collection of resting-state and walking EEG windows. With validation in larger, independent, and more balanced clinical cohorts, the proposed framework could support objective EEG-based decision support and provide a foundation for portable or wearable systems for longitudinal monitoring of PD-related cortical changes in clinical and home-based settings.