Medical Imaging

Proposed a GAN-based desmoking framework enhanced with a refined dark channel mask as an embedded guide, significantly improving laparoscopic video clarity without requiring paired data.

Achieved superior performance in both quantitative metrics (PSNR: 32.6, SSIM: 0.92) and referenceless scores (JNBM, FADE, REA), outperforming multiple existing smoke removal models on the LVQ dataset.

Validated the method through a user study with 25 surgeons, where the model’s outputs were most preferred for clinical usability and aesthetic quality, demonstrating strong potential for surgical integration.

Proposed a CycleGAN-based framework for laparoscopic video desmoking, enhanced with inter-channel (IC) and dark channel prior (DC) loss functions to minimize residual smoke and restore image clarity in unpaired training scenarios.

Achieved significant improvements in PSNR (38.6) and SSIM (0.94) over existing state-of-the-art methods on the LVQ dataset, highlighting strong performance in high-smoke environments.

Validated the model through a user study with 25 surgeons, confirming that the desmoked outputs were consistently preferred for both visual quality and surgical usability—indicating strong potential for real-world clinical integration.

Conducted a systematic comparison of data-driven (computational) and clinician-guided feature selection techniques—including Chi-Square, Mutual Information, Sequential Forward/Backward Selection, and Medically Important Features—for predicting heart disease.

Demonstrated that computational methods (especially SFS and MI) consistently outperformed expert-selected features in accuracy and F1 score, with SVM and Random Forest classifiers achieving >85% accuracy using optimized subsets.

Provided insights into optimal feature subset sizes for various ML algorithms, validating that reduced-dimensional, interpretable models can still maintain high predictive performance for real-world, resource-constrained healthcare scenarios.

Developed a visualization framework integrating eye-tracking and capsule endoscopy, enabling pixel-level analysis of gaze patterns to study diagnostic behavior and reduce perceptual errors in medical imaging.

Conducted experiments with 536 annotated endoscopy images, revealing that expert search strategies commonly follow circular or radial patterns, enhancing interpretability and coverage.

Demonstrated effective gaze calibration and pixel error correction techniques, producing accurate heatmaps and saliency maps that can inform both radiology education and intelligent diagnostic systems.

Developed a custom eye-tracking and visualization system integrated with capsule endoscopy image analysis, capturing expert gaze behavior across 536 annotated medical images from two datasets.

Identified consistent circular and radial visual search patterns through heatmaps, saccade/area coverage analysis, and fixation metrics—offering insight into expert diagnostic strategies and cognitive load.

Demonstrated the potential of this dataset and framework to inform radiology education, support automated diagnostic tools, and enhance future deep learning integration for medical image interpretation.

Developed a GAN-based supervised image-to-image translation framework to effectively eliminate surgical smoke from laparoscopic videos, improving visibility during procedures.

Enhanced the Perceptual Adversarial Network (PAN) with modified SSIM and multi-scale loss functions to improve visual realism and clarity, outperforming traditional physical and classical de-smoking methods.

Validated the model using the LVQ dataset and a user study with 25 surgeons, who consistently preferred the proposed method for both aesthetic quality and surgical usability.