AI for Healthcare - Syllabus
Technologies Enabling AI
• Machine Learning
• Deep Learning
• NLP
• Computer Vision
• Robotics
Types of Learning
• Supervised Learning – Given training data + desired outputs
(labels)
• Regression
• Classification
• Unsupervised Learning – Given training data (without desired
outputs)
• Clustering
• Density estimation
• Semi-supervised Learning – Given training data + few desired
outputs
• Reinforcement Learning
• Rewards from sequence of action
Deep Learning – ANN – Similar to Biological neuron
Biological neuron:
Fundamental processing
elements of a neural network is a
neuron
[Link] inputs from other
source
[Link] them in someway
3. Performs a generally nonlinear
operation on the result
[Link] the final result
Synapses - Basis of learning and memory
Activation
Summing function
function
CNN (Convolution Neural Network)
Class of Deep learning (DL)
Similar to ANN with learnable parameters, such as, weights, bias, etc.
3 basic components of CNN are:
•Convolution layer
•Pooling layer
•Output layer (Fully connected layer)
Convolution and Pooling Layer
Computers read images as pixels (N*N*3)
The convolution layers uses the convolution filter to detect presence of specific
features or patterns in original image
Compresses the image into smaller dimension with same depth as input image
(M*M*3)
This filter is then convolved (slided) across the width and height of the input image,
and a dot product is computed to give the new image.
Max pooling layer
Medical Image analysis
• Extracting meaningful information from medical images
• Manual Vs. Automated image analysis
• Manual: Human expert-driven (radiologists, pathologists)
• Automated: AI or algorithm-driven systems
• Goals: Improved diagnostic accuracy, speed and scalability in
• Detection (e.g.: tumors, fractures)
• Quantification (e.g.: size, volume)
• Localization (e.g.: tumor boundary)
Manual Image analysis
• Performed by trained medical professionals
• It involves:
• Visual inspection
• Measurement tools (e.g., rulers, annotations)
• Experience-based interpretation
• Strengths:
• Clinical context awareness
• Intuition and judgment
• Limitations:
• Time-consuming
• Prone to fatigue and inter-observer variability
• Not scalable for high-volume analysis
Automated Image analysis
• Uses AI models (ML/DL) to extract, classify, and segment features
• It involves:
• Tumour classification
• Organ/tissue segmentation
• Anomaly detection
• Benefits:
• Fast and consistent
• Intuition and judgment
• Scalable and deployable across institutions
• Limitations:
• Requires annotated data
• May lack clinical interpretability
• Needs validation in real-world settings
Overview of AI techniques in medical image
analysis
• Machine Learning (ML):
• Traditional features + classifiers (SVM, Random Forest)
• Deep Learning (DL):
• Convolutional Neural Networks (CNNs)
• Transfer learning (U-Net, ResNet, etc.)
• Example: Predict if a chest X-ray indicates COVID-19 or not
• Workflow: Input image → Pre-processing → CNN → Softmax classifier
Case study 1: COVID-19 Diagnosis from Chest X-rays
• Data collection:
• Labeled chest X-ray images (COVID+, COVID−)
• Public datasets: COVIDx, NIH ChestX-ray14, Kaggle
• Pre-processing:
• Resize, normalize, denoise
• Data augmentation (flip, rotate)
Case study 1: COVID-19 Diagnosis from Chest X-rays
• Pre-processing:
• Resize, normalize, denoise
• Data augmentation (HE, CLAHE)
• Model training:
• CNN-based architectures
• (e.g., EfficientNet, VGG, DenseNet)
• Prediction and evaluation:
• Outputs probability of COVID-19
• Metrics: Accuracy (>90%), Precision, Recall
Case study 1: AI-assisted Diagnosis of COVID-19
• AI solution for screening COVID-19 patients
based on conventional chest/lung X-rays
• The system uses Principal Component Analysis
(PCA) and clustering techniques to distinguish
between L and H phenotypes of COVID-19
• AI tool will help doctor to make better
healthcare decision such as:
• Patient would need mechanical ventilator to survive
(H phenotype)
• Patient will have lung injury if ventilator is applied
(L phenotype)
Case study 2: AI-assisted Breast cancer detection
• Regular screening and early detection - improves breast cancer survival rate – higher
toll on radiologists.
• AI Tools can process images faster and provide valuable insights for decision-making,
empowering radiologists to make evaluations up to 15-20% faster.
• A recent collaboration between Google Health and Physicians yielded very promising
results.
• The group is working on a supporting tool for mammography readings. When put to
the test, the AI algorithm was as accurate as a serial reading done by two doctors.
• It was the speed, however, that was a significant differentiator—the system reduced
the workload of the second reader by 88%.
Case study 3: AI for Digital Pathology
• AI can assist pathologies with:
• Image analysis and interpretation
• Detailed inspections of sample tissues
• Pathology types matching to earlier cases
• Diagnosis accuracy and early detection
• A group of cancer researchers recently analyzed a public database of
• WSIs from 11,000 cancer patients, featuring 32 cancer subtypes.
• The tool identified different types of pathology on frozen section slides with high accuracy:
• 93% for bladder urothelial carcinoma
• 97% for kidney renal clear cell carcinoma
• 99% for ovarian serous cystadenocarcinoma
• And performed equally accurately with histopathology slides:
• 98% for prostate adenocarcinoma
• 99% for skin cutaneous melanoma
• 100% for thymoma
Generative AI for Medical Imaging
• Generative AI = AI models that can create new, realistic data
• Common model: GANs, VAEs, Diffusion Model
• Role in medical imaging: enhance, synthesize, and analyze
• Main applications of Gen AI in medical imaging are:
• Data augmentation for Training
• Image reconstruction and enhancement
• Segmentation assistance
• Image modality translation
• Disease progression simulation
• Anomaly detection
Data Augmentation for Training
• Generates synthetic images to enlarge datasets
• Helps balance rare disease classes
• Reduces model over fitting
• Example: Original CT image + Synthetic CT image
Image reconstruction and Enhancement
• Improves quality of noisy/low-resolution scans
• Useful in MRI, CT when scan time/dose is limited
• AI fills missing data → faster, safer imaging
• Example: Low resolution MRI Gen AI High resolution MRI
Segmentation assistance and Modality translation
Segmentation Assistance:
• AI models generate accurate organ/tumor boundaries
• Enhances radiologist efficiency
• Useful in tumor detection, organ delineation
Modality translation:
• Converts one imaging type to another (e.g., MRI → CT)
• Reduces need for multiple scans
• Saves cost and reduces radiation exposure
• Example: MRI image CT image
Disease Progression simulation and Anomaly detection
Disease Progression Simulation:
• Generative models predict how disease evolves
• Helps doctors visualize tumor growth or recovery.
• Supports treatment planning
Anomaly Detection:
• Generative models learn "normal anatomy"
• Deviations indicate potential disease
• Useful for early detection in chest X-rays, brain MRI
Challenges and Future scope
Challenges:
• Data privacy & security
• Regulatory approval of synthetic images
• Bias in training data
Future scope:
• AI-assisted radiology
• Robust clinical adoption
