Course Information
Instructor
Matt Ebrahim (Mahmoud Ebrahimkhani)
Email: m.ebrahimkhani@northeastern.edu
Office Hours: Saturdays 11:00 – 12:00 PM ET (Zoom)
Course Details
Meeting Times: Saturdays 12:30 – 3:00 PM ET
Format: Virtual (Zoom)
Course Description
This course provides a comprehensive introduction to generative AI techniques and their applications in healthcare. Students will explore cutting-edge methods spanning from foundational deep learning architectures to state-of-the-art generative models, with hands-on implementation experience throughout.
Course Schedule
Session 1: Introduction to Deep Learning and Generative AI in Healthcare
January 10, 2026
Course overview and objectives; transformative impact of deep learning across clinical medicine; applications in medical image diagnosis (radiology, pathology, ophthalmology); protein structure prediction with AlphaFold; AI-driven drug discovery pipelines; unique challenges of biomedical data (class imbalance, annotation scarcity, privacy constraints, regulatory requirements)
Session 2: Deep Learning Fundamentals
January 17, 2026
Perceptrons and multilayer feedforward networks; forward and backward propagation of gradients; activation functions (ReLU, sigmoid, tanh); optimization algorithms (SGD, Adam, RMSProp); learning rate scheduling; regularization techniques (dropout, batch normalization, weight decay); bias-variance tradeoff and overfitting in medical settings
Session 3: Convolutional Neural Networks
January 24, 2026
Convolution operations, stride, padding, and receptive fields; pooling layers for spatial downsampling; LeNet, AlexNet, VGG, and ResNet architectures; skip connections and residual learning; feature maps and hierarchical representation learning; image classification pipelines for biomedical imagery
Session 4: Medical Image Classification
January 31, 2026
Transfer learning from ImageNet to medical domains; feature extraction vs. fine-tuning strategies; pretrained architectures (ResNet, DenseNet, EfficientNet); MedMNIST and clinical imaging datasets; end-to-end classification pipelines; evaluation metrics (accuracy, AUC-ROC, F1-score, confusion matrix); handling class imbalance with weighted loss and oversampling
Session 5: Transfer Learning and Ensemble Learning
February 7, 2026
Advanced fine-tuning strategies (layer freezing, differential learning rates, gradual unfreezing); ensemble learning theory; soft voting (probability averaging) and hard voting (majority vote) ensembles; bagging and model diversity; combining heterogeneous architectures; performance gains and computational trade-offs of ensembles on medical imaging benchmarks
Session 6: Medical Image Segmentation
February 14, 2026
Semantic vs. instance segmentation; U-Net encoder-decoder architecture with skip connections; Attention U-Net and 3D volumetric extensions; combined Dice and cross-entropy loss; evaluation metrics (Dice coefficient, IoU, Hausdorff distance); clinical applications in tumor delineation, organ segmentation, and surgical planning; data augmentation strategies for limited medical annotations
Session 7: NLP in Medicine I
February 21, 2026
Clinical text preprocessing and tokenization; named entity recognition for diseases, medications, and procedures; word embeddings (Word2Vec, GloVe) and contextual representations; transformer architecture (self-attention, multi-head attention, positional encoding); BERT and BioBERT for clinical NLP; information extraction from electronic health records (EHRs); ICD coding and clinical concept normalization
Session 8: NLP in Medicine II
February 28, 2026
Recurrent neural networks and LSTMs for sequential clinical data; sequence-to-sequence models with encoder-decoder attention; character-level and subword language models; clinical note summarization and discharge summary generation; question answering over medical literature; ROUGE and BLEU evaluation metrics; GPT-style models and instruction tuning for healthcare text tasks
Session 9: Generative Models — Autoencoders, VAEs, and GANs
March 14, 2026
Autoencoder architecture and latent space properties, variational autoencoders and probabilistic encoding, reparameterization trick, ELBO loss, GAN generator/discriminator training dynamics, Wasserstein GANs, mode collapse and vanishing gradients, medical imaging applications (MRI reconstruction, MR-to-CT synthesis, cycle-consistent GANs)
Session 10: GANs for Medical Image Synthesis — MR-to-CT and MRI Acceleration
March 21, 2026
Deep dive into GAN architectures for clinical imaging: U-Net generator with skip connections, PatchGAN discriminator, multi-component loss (pixel-wise MSE, frequency-domain MSE, VGG perceptual loss, adversarial loss), MR-to-CT synthesis using cycle-consistent GANs for unpaired image translation, MRI acceleration with DAGAN (deep learning augmented GAN) using k-space undersampling and refinement learning
Session 11: Denoising Diffusion Probabilistic Models for Medical Image Generation
March 28, 2026
Denoising Diffusion Probabilistic Models (DDPMs): forward diffusion process, reverse denoising process, noise scheduling, sinusoidal time embeddings; U-Net with ResNet blocks and self-attention for the denoising network; CycleGAN with residual learning for multi-modality synthesis (MRI-to-CT, 3T-to-7T MRI); text-to-image generation (DALL-E 2, Imagen); diffusion models for medical imaging tasks including segmentation, reconstruction, anomaly detection, and histopathology image generation; genotype-conditioned synthesis for glioma classification; evaluation metrics (FID, Inception Score, Improved Precision/Recall)
Assignments
Homework 1: Transfer Learning and Ensemble Methods
Train pretrained models on BloodMNIST dataset, implement ensemble methods (averaging and voting), and compare performance across different architectures.
Homework 2: Medical Image Segmentation and NLP
Implement a 3D Attention U-Net for volumetric medical image segmentation, train a character-level LSTM for medical text generation, and build an RNN with additive attention for clinical note classification.
Homework 3: Variational Autoencoders for Medical Image Generation
Implement a Convolutional VAE from scratch — encoder, reparameterization trick, decoder, and ELBO loss — and apply it to PathMNIST colon pathology slides. Evaluate generation quality via reconstructions, random sampling, and latent space interpolation.
Resources
Required Textbooks
- Bishop, C. M., & Bishop, H. (2024). Deep Learning: Foundations and Concepts. Springer. [PDF]
- Chollet, F. (2021). Deep Learning with Python (2nd ed.). Manning.
- Tunstall, L., von Werra, L., & Wolf, T. (2022). Natural Language Processing with Transformers. O'Reilly.
Development Tools
- PyTorch – Primary deep learning framework
- Hugging Face – Transformers and model hub
- Jupyter Notebook – Interactive coding environment
- Google Colab – Free GPU access
Additional Materials
Grading
- Homework Assignments (60%): Four assignments, 15% each
- Final Project (30%): Team project with proposal, implementation, and presentation
- Participation (10%): Class attendance and engagement