Goal

• Detect and classify lumbar degenarative spine (LSD) conditions using lumbar sping MR images
• More information, credits, and data are on the Kaggle competition website.

The 5 conditions are

• Left Neural Foraminal Narrowing
• Right Neural Foraminal Narrowing
• Left Subarticular Stenosis
• Right Subarticular Stenosis
• Spinal Canal Stenosis

For each condition, classify the 5 intervertebral disc levels (IDL) = location of the compression

• L1/L2
• L2/L3
• L3/L4
• L4/L5
• L5/S1

Classify how severe of compression of the spinal cord due to foramina/subarticular exit (severity score/level) weights

• Normal/mild = 1
• Moderate = 2
• Severe = 4

So, essentially, for each observation, we have 5*5=25 output variables, each of which needs to be classified as one of the 3 severity levels.

Steps:

1. Conduct EDA to understand the data and distributions: rsna-lumbar-eda.ipynb
2. Develop essential functions to deal with data like making metadata, showing images, and coordinating directories: rsna-lumbar-keyfunctions.ipynb
3. Examine the coordination of pathology and look into several images given the previously built functions: rsna-lumbar-imageprocessing-and-coordpathology.ipynb
4. Convert the DICOM images into PNG for convenience and select subsets of the images for training: rsna-making-dataset-png.ipynb
5. Make Dataset, DataLoader, transformations, define a ResNet50 model and training loop on AdamW optimizer and cuda GPU T4x2: rsna-training-resnet50-gpu.ipynb. I will add more details to this. I had such a great time to learn from other notebooks on how to handle the data/image processing problem that involves multiple outcome variables for 3-class classification.

Some visual info:

• Left Neural Foraminal Narrowing
• Right Neural Foraminal Narrowing

5 intervertebral disc levels	Severity

• Left Subarticular Stenosis
• Right Subarticular Stenosis

5 intervertebral disc levels	Severity

• Spinal Canal Stenosis

5 intervertebral disc levels	Severity