Machine Learning Fundamentals

ML Fundamentals - AI's Learning Curve

• Machine Learning (ML): AI subset; systems learn from data.
• Types:
  • Supervised: Labeled data (e.g., image classification: tumor/normal).
  • Unsupervised: Unlabeled data (e.g., anomaly detection).
  • Reinforcement: Rewards/penalties (less common in radiology).
• Key Terms:
  • Features: Input variables (pixel values).
  • Labels: Output variables (disease state).
  • Model: Learned mathematical representation.
  • Training: Learning process.
  • Inference: Prediction on new data.

⭐ Arthur Samuel coined "Machine Learning" in 1959 (IBM), known for computer checkers.

Supervised Learning - Guided Predictions

• Core Concept: Learning from labeled data, mapping inputs to outputs ($X \rightarrow Y$).
• Primary Tasks:
  • Classification: Assigns inputs to predefined categories. E.g., classifying a lung nodule as benign or malignant.
  • Regression: Predicts continuous numerical values. E.g., estimating tumor volume or bone age.
• Common Algorithms:
  • Linear/Logistic Regression: Models relationships to predict outcomes.
  • Support Vector Machines (SVM): Finds an optimal boundary to separate data classes.
  • Decision Trees: Uses a tree-like structure for decision-making.
  • k-Nearest Neighbors (k-NN): Classifies based on the 'k' closest labeled examples.
• Radiology Applications:
  • CADe (Computer-Aided Detection): Highlights suspicious regions for radiologists (e.g., microcalcifications).
  • CADx (Computer-Aided Diagnosis): Characterizes lesions to assess malignancy risk.

⭐ The vast majority of currently FDA-approved AI algorithms in diagnostic radiology are based on supervised learning techniques.

Unsupervised Learning & Data Prep - Finding Hidden Patterns

• Core: Discovers patterns in unlabeled data.
• Tasks:
  • Clustering: Groups similar data. E.g., k-Means (partitions data into $k$ clusters by similarity).
  • Dimensionality Reduction: Reduces features. E.g., PCA (finds key patterns, reduces complexity/noise).
• Data Prep (GIGO: Garbage In, Garbage Out):
  • Cleaning: Handle missing values, outliers.
  • Normalization/Standardization: Scale features.
  • Augmentation: Expand image data (rotate, scale). Vital for medical images.
• Dataset Splitting (70% Train / 15% Val / 15% Test):
  • Train: Model learning.
  • Validate: Tune hyperparameters; prevent overfitting.
  • Test: Unbiased final performance evaluation.

⭐ Data augmentation is a critical technique in medical imaging AI to artificially expand limited datasets and improve model generalization, thereby reducing overfitting.

Model Metrics & CNNs - AI's Checkup

• Confusion Matrix: Elements: TP, TN, FP, FN.

  • TP: Correct positive; TN: Correct negative.
  • FP: False alarm (Type I error); FN: Missed detection (Type II error).

• Key Metrics:

  • Sensitivity (Recall): $S_n = TP/(TP+FN)$ (detects disease).
  • Specificity: $S_p = TN/(TN+FP)$ (confirms absence).
  • Precision (PPV): $PPV = TP/(TP+FP)$ (true positives among positive calls).
  • Accuracy: $Acc = (TP+TN)/(TP+TN+FP+FN)$ (overall correct).

• AUC-ROC: Area Under ROC Curve. Plots $S_n$ vs (1-$S_p$). Higher AUC = better discrimination.

• Common Pitfalls:

  • Overfitting: Good on training data, poor on test data.
  • Underfitting: Poor on both training and test data.

• Deep Learning (DL): ML using deep Artificial Neural Networks (ANNs).

• Convolutional Neural Networks (CNNs): DL for image analysis; hierarchical feature learning.

  • Key Layers: Convolutional (features), Pooling (reduce dimensionality), Fully Connected (classify).

⭐ The Area Under the ROC Curve (AUC-ROC) is a widely used metric to evaluate the performance of binary classification models in radiology, representing the trade-off between sensitivity and specificity across different thresholds.

High‑Yield Points - ⚡ Biggest Takeaways

• Supervised learning (e.g., CNNs) is vital for image classification and segmentation tasks.
• Unsupervised learning aids in pattern discovery and anomaly detection in large datasets.
• Deep Learning, particularly Convolutional Neural Networks (CNNs), are foundational for radiological image analysis.
• Training, validation, and test datasets are essential to prevent overfitting and ensure generalizability.
• Key performance metrics include AUC-ROC, sensitivity, specificity, and accuracy.
• Awareness of algorithmic bias and the need for explainable AI (XAI) is growing in radiology AI applications.