Medical AI & Imaging
Work with medical images and clinical data: DICOM, segmentation with U-Net, MONAI, and nnU-Net, pathology, foundation models, bias, and regulation.
Before this: Deep Learning. Computer Vision helps for the segmentation steps. No medical background needed.
-
1h 11m
Step 1What makes healthcare data different
MIT lecture on clinical data, its messiness, and its stakes.
1. What Makes Healthcare Unique?MIT OpenCourseWare
-
11 min
Step 2How ML reads medical images
JAMA explainer on training image models against expert labels.
Machine Learning For Medical Image Analysis - How It WorksJAMA Network
-
22 min
Step 3Read DICOM files in Python
Load scans and their metadata with pydicom.
1. Dicom In PythonArbois Code Media
pydicom/pydicom★ 2.2k -
12 min
Step 4Sensitivity and specificity
The two error rates clinicians ask about first.
Machine Learning Fundamentals: Sensitivity and SpecificityStatQuest with Josh Starmer
-
32 min
Step 5U-Net for segmentation
The encoder-decoder with skip connections behind most medical segmentation.
U-Net clearly explained | Image Segmentation with AITileStats
-
37 min
Step 6Segment organs in 3D CT
Train a 3D segmentation model on CT scans in Python.
Python AI Organ Segmentation TutorialMr. P Solver
Project-MONAI/MONAI★ 8.9k -
49 min
Step 7nnU-Net
A U-Net that configures its own preprocessing and training per dataset.
nnU-Net: Automated Design of Deep Learning Methods for Biomedical Image Segmentation | Paul JägerIntelligent Medical Systems
MIC-DKFZ/nnUNet★ 9.1k -
14 min
Step 8Segment anything in medical images
MedSAM adapts Segment Anything to prompt-based medical segmentation.
MedSAM Segment Anything in Medical Images Universal medical image segmentation.AI WITH Rithesh
bowang-lab/MedSAM★ 4.5k -
56 min
Step 9Machine learning for pathology
MIT lecture on learning from huge whole-slide tissue images.
12. Machine Learning for PathologyMIT OpenCourseWare
-
47 min
Step 10Foundation models in radiology
Stanford researchers on large pretrained models for medical images and reports.
Foundation Models in RadiologyMIDRC Media
-
35 min
Step 11Bias in health AI
Where bias enters clinical models, from labels and proxies to deployment.
Where Are We Now? Bias in Health AIUniversity of California Television (UCTV)
-
19 min
Step 12How AI medical devices get cleared
How the FDA regulates diagnostic imaging AI software in the US.
David Giese's Presentation on FDA Regulation of Diagnostic Imaging AIInnolitics
Build it
Watching isn't learning. Pick one and put it on GitHub.
- Load a public CT series with pydicom, window it, and save an axial, coronal, and sagittal view.
- Train a 2D U-Net on a public segmentation dataset and report Dice score per case.
- Train a chest X-ray classifier and report sensitivity, specificity, and AUC on a held-out hospital split if available.