• I’m a machine learning scientist/engineer with extensive experience in applications of deep learning. I have mainly worked on supervised and unsupervised information retrieval, image segmentation and classification tasks. I implemented and evaluated several ML models in Python (Tensorflow, PyTorch, Sklearn).

Email / LinkedIn / Google Scholar

Machine Learning Researcher @ Vector Institute (Sep 2021 - Jan 2022)

Semnatic Segmentation and object detection in medical images

• Built an end-to-end AI recommendation system for x-ray image segmentation to enhance radiology diagnoses with machine learning
• Achieved 0.74 accuracy
• Trained UNet++ with EfficientNet-B4 backbone, two stages of training on public and private datasets
• Data augmentations: horizontal flip, rotation, random gamma, etc
• Loss function: combination of Dice loss and binary cross-entropy
• Training setups: Adam optimizer, Cosine Annealing scheduling, early validation stopping, and K-fold cross-validation
• Post-Processed images to to draw attention to suspicious areas
• Coded in Python using Tensorflow, OpenCV, Scikit-Learn, Numpy, Pandas, etc.

Graduate Student Research Assistant @ Kimia Lab (Sep 2019 - Sep 2021)

Deep learning application in medical image analysis

• Proposed multi-magnification CBIR using features derived by deep learning
• Implemented unsupervised hashing algorithms to increase information retrieval efficiency and statistical methods for deep feature aggregation
• Up to 11% accuracy improvement
• 
  
  

-Binary convolutional neural networks for medical image classification; Implemented rotated binary network and evaluation measures in Python; Improved the inference speed by 10 times and memory storage by 21 times compared to the full-precision model.

• Implemented U-Net topologies for artifact removal and find ROI in pathology images
• Achieved 0.99 accuracy: Jaccard Index=0.95, Dice Score: 0.97, Sensitivity: 0.99, Specificity: 0.99

• Developed deep feature extractor for medical image representation using transfer learning. Improved image search accuracy by 30% compared to pre-trained DenseNet