Yasser El Jarida

PhD Student, Medical AI

UM6P College of Computing, Ben Guerir, Morocco

Education

PhD in Computer Science (Medical AI) — UM6P College of Computing, Ben Guerir (2024–Present)
Computer Engineering: Big Data and Cloud Computing — ENSET, Mohammedia (2021–2024)
Preparatory Classes in Mathematics and Physics (CPGE) — Ibn Abdoun High School, Khouribga (2019–2021)
Baccalaureate in Mathematical Sciences A — Abou El Kacem Ezzayani High School, Khenifra (2018–2019)

Keywords: Transformers, Cross-Attention, PyTorch, ECG Analysis, Surgical Phase Detection, Synthetic Data

Conducting advanced research on AI for healthcare, focusing on cardiology (ECG analysis) and surgical workflow optimization via intelligent video analysis.
Designed and implemented a Boundary-Aware FACT (Frame–Action Cross-Attention Temporal) model for surgical phase segmentation, achieving 2nd place at the MICCAI 2025 OMNIA SICS155 Challenge.
Introduced a lightweight boundary supervision head and a boundary-weighted temporal smoothing loss, improving segmentation accuracy by +1.3% and F1-score by +1.5 without changing inference cost.
Explored transformer-based architectures (FACT, SurgFormer, TimeSformer, VideoMAE‑v2) for temporal understanding and phase recognition in surgical videos.
Collaborated with clinical experts to ensure interpretability, efficiency, and workflow alignment of deployed models.
Prior work: synthetic datasets and CNN regression (ResNet50, EfficientNet‑B0, InceptionV3) for instant particle size distribution (PSD); published at CVPR 2025 SynData4CV.

Keywords: Python, YOLOv8, SAM, CVAT, DeepFill v2, ResNet50, Streamlit

Built a CV pipeline with YOLOv8 + SAM for detection, segmentation, and reflectivity assessment of CSP mirrors.
Improved data quality with image inpainting (DeepFill v2), achieving R² = 0.94 for reflectivity prediction (ResNet50).
Deployed a Streamlit dashboard for interactive evaluation and visualization.

Keywords: Python, TensorFlow, CNNs, YOLOv8

Implemented vision-based violence detection using YOLOv8 with extensive tuning and validation.
Analyzed precision, recall, and F1 to guide iterative improvements.

Instant Particle Size Distribution Measurement Using CNNs Trained on Synthetic Data. CVPR 2025 Workshop: SynData4CV (Accepted), 2025.

Task: multi‑phase recognition on SICS‑155 (155 videos, 19 phases).
Model: FACT (Frame–Action Cross‑Attention) with I3D features.
Boundary head: lightweight Conv1D predicts per‑frame transition probability p_b(t), used only during training.
Losses: boundary BCE + boundary‑weighted total variation (gated by (1 − p_b(t))^γ) plus standard frame/token/attention losses.
Result: cleaner phase transitions, fewer spurious short segments, +1.3% accuracy and +1.5 F1 vs. baseline; 2nd on leaderboard.

Blender‑generated synthetic dataset spanning shapes, textures, and lighting.
CNN regressors (ResNet50, EfficientNet‑B0, InceptionV3) predict PSD metrics (d10/d50/d90) with strong efficiency/accuracy trade‑offs.

Vision Transformers + CNN backbones for robust ECG representation learning.
Focus: noise robustness, rhythm irregularity detection, and interpretability for clinical reliability.

Human Research: Data or Specimens Only Research (Basic Course) — CITI Program / MIT Affiliates
Fundamentals of Accelerated Computing with CUDA Python — NVIDIA
Neural Networks and Deep Learning — DeepLearning.ai
Supervised Machine Learning: Regression and Classification — DeepLearning.ai

2nd Place — MICCAI 2025 OMNIA SICS155 Surgical Phase Recognition Challenge, 2025
Accepted Paper — CVPR 2025 Workshop (SynData4CV), 2025
1st Place — Hackathon: Blockchain and AI at the Service of Health, 2023