Research
PhD Thesis
Automatic Detection of Pathologies in Conventional Scanner Data
This thesis investigates AI-based detection and characterization of pulmonary pathologies in conventional CT scans, with a focus on reducing annotation requirements. It explores supervised, weakly supervised, and unsupervised learning strategies to detect anomalies and support radiologists in identifying incidental findings beyond the initial clinical indication.
Author: Aissam Djahnine · Supervisor: Loïc Boussel
Projects
Segmentation and Measurement of Skeletal Muscle Areas on CT Scans
Developed deep learning-based segmentation at L3/L2/L1 levels to improve sarcopenia estimation with stronger data augmentation and post-processing for cleaner masks.
Cell Image Segmentation with CycleGAN
Built a cycle-consistent GAN workflow for cell image segmentation with limited labels, enabling synthetic data generation and context-aware training.
Histopathological Image Generation Using GANs
Implemented GAN-based synthetic histopathology generation to address small datasets and improve downstream detection/segmentation/tracking tasks.
Color Transfer in Correlated RGB Space
Implemented correlated color transfer to adapt visual tone and style between images, based on the method by Xuezhong Xiao and Lizhuang.
Search Algorithms (BFS, DFS, Dijkstra, A*)
Implemented core search strategies and heuristics (Manhattan/Euclidean) to provide practical foundations for AI pathfinding and problem-solving.
Speech Denoising via Spectral Subtraction
Applied spectral subtraction for acoustic noise suppression in speech signals, improving clarity and intelligibility in noisy environments.
Markov Decision Process
Explored decision-making under uncertainty with MDP modeling and dynamic programming principles for optimization across applied domains.
Metro Problem (Dijkstra in Paris Subway)
Built a shortest-path tool over subway graph data with user-friendly station matching and efficient path computation in C++.