Biomedical AI Projects

🌐 Overview

This repository presents two AI-driven biomedical projects developed during my exchange semester in POLIMI:

1. Automatic Segmentation of Vocal Tract Articulators using dsMRI and IMU-Net
2. Lung Cancer CT Scan Classification with Confidence Estimation

Each project focuses on solving a unique challenge in medical imaging through deep learning, emphasizing robustness, interpretability, and clinical relevance.

---

🔜 Project 1: Dynamic MRI Vocal Tract Segmentation

Objective

To segment vocal tract articulators from dynamic speech MRI (dsMRI) contaminated by Salt & Pepper noise using the IMU-Net architecture.

Highlights

• Architecture: Modified U-Net with residual connections, dilated convolutions, and dropout for robustness
• Preprocessing: Normalization, 3x3 Median filtering
• Custom Loss: Dice + Focal loss for precise segmentation across 7 binary masks (e.g., tongue, soft palate)
• Validation: k-Fold cross-validation on 820 frames
• Evaluation Metric: Mean DICE Score

Applications

• Diagnosis of speech disorders (apraxia, dysarthria)
• Functional articulation analysis
• Therapy monitoring, surgical planning, prosthetic design
• Research on sleep/breathing disorders

Real-World Testing

• Tested on dsMRI of a control subject saying "microscopic"
• Promising results; not yet validated on pathological cases

---

🌐 Project 2: Lung CT Scan Classification

Objective

To classify lung nodules from CT scans as benign or malignant with confidence estimation using deep learning models.

Highlights

• Models: Baseline CNNs, VGG19, MobileNetV2, EfficientNetB0, ResNet50
• Augmentation: Applied before/after split; class weights used for balancing
• Explainability: Grad-CAM applied to MobileNetV2
• Performance: Best results with VGG19 on zoomed slices (Accuracy: 89%, F1: 0.89)

Key Takeaways

• Zoomed slices outperformed full slices due to focused information
• Model generalization highly dependent on data imbalance strategies
• Grad-CAM shows potential for better diagnostic trust

Note

NB: Les résultats ont été obtenus sur un jeu de données fourni par notre professeur, dont le contenu nous était inconnu au départ.

---

🔬 Technologies Used

• Frameworks: TensorFlow, Keras
• Techniques: Transfer learning, data augmentation, k-fold CV, Grad-CAM, custom loss functions
• Evaluation: Accuracy, Precision, Recall, F1, Mean DICE, class-specific metrics

---