---

Overview

The midterm project is your first opportunity to apply deep learning to a medical imaging problem end-to-end. You’ll work on a self-selected project from the approved options below, building on skills from Modules 1-5.

Deliverables:

1. Working code (Jupyter notebook or Python script)
2. Mini field guide (2-3 pages)
3. Brief presentation (5-7 minutes)

---

Project Options

Choose ONE of the following tracks:

Option A: Chest X-Ray Classification

Task: Build a classifier to detect one or more conditions from chest X-rays.

Suggested Datasets:

• NIH ChestX-ray14 — 112k images, 14 disease labels
• CheXpert — 224k images from Stanford
• RSNA Pneumonia Detection — Kaggle challenge data

Minimum Requirements:

• Binary or multi-label classification
• Train/val/test split with proper evaluation
• At least one baseline and one CNN model
• ROC curve and confusion matrix

---

Option B: Medical Image Segmentation

Task: Segment anatomical structures or lesions from medical images.

Suggested Datasets:

• ISIC Skin Lesion — Dermoscopy images with lesion masks
• LiTS Liver Tumor — CT liver/tumor segmentation
• DRIVE Retinal Vessels — Fundus images with vessel masks
• Any MONAI dataset from the Model Zoo

Minimum Requirements:

• U-Net or similar architecture
• Dice score evaluation
• Visualization of predictions vs ground truth
• At least 3 example cases in your report

---

Option C: Retinal Image Analysis

Task: Classify diabetic retinopathy severity or detect other retinal conditions.

Suggested Datasets:

• APTOS 2019 Blindness Detection — Kaggle DR classification
• Messidor-2 — DR grading dataset
• ODIR-5K — Multi-label ocular disease

Minimum Requirements:

• Handle class imbalance (DR severity is very imbalanced)
• Transfer learning from pretrained model
• Quadratic weighted kappa evaluation (standard for DR)
• Error analysis on misclassified cases

---

Option D: Your Own Project (Requires Approval)

Have a different medical imaging project in mind? Propose it!

To get approval:

1. Email the instructor by March 20 with:
   • Dataset description and access plan
   • Clinical problem and why it matters
   • Proposed approach
2. Wait for approval before starting

---

Requirements

Code (50 points)

Component	Points
Data loading and preprocessing	10
Model architecture (appropriate for task)	10
Training pipeline (loss, optimizer, logging)	10
Evaluation metrics (appropriate for task)	10
Code quality (readable, documented)	10

Mini Field Guide (35 points)

Write a 2-3 page document covering:

Section	Points
Problem Statement — What clinical problem does this address?	5
Data Description — What data did you use? Limitations?	5
Methods — What approach did you take and why?	10
Results — Key performance metrics with interpretation	10
Limitations & Next Steps — What would you do differently?	5

Presentation (15 points)

5-7 minute presentation covering:

• The problem and why it matters
• Your approach
• Key results
• One thing you learned

Presentations will be during class on April 1.

---

Submission

Submit via GitHub (link TBD):

1. code/ — Your notebooks and/or scripts
2. field_guide.pdf — Your mini field guide
3. slides.pdf — Presentation slides
4. README.md — How to run your code

---

Timeline

Date	Milestone
Mar 18	Project released, choose your track
Mar 20	Deadline for custom project proposals
Mar 25	Recommended: Data loaded, baseline running
Mar 31	In-class work session
Apr 1	Presentations in class
Apr 1	Code + Field Guide due by 11:59 PM

---

Tips

• Start with the data — Make sure you can load and visualize before building models
• Get a baseline working first — A simple model that runs beats a complex model that doesn’t
• Document as you go — Don’t leave the field guide until the last minute
• It’s okay to fail — If your model doesn’t work well, analyze why. That’s valuable!
• Use what you learned — HW2 (DICOM), HW5 (deep learning) are directly relevant

---

Grading Rubric Summary

Component	Points
Code	50
Mini Field Guide	35
Presentation	15
Total	100

---

Resources

• MONAI Tutorials
• PyTorch Image Classification Tutorial
• Papers With Code - Medical Imaging
• Office hours — Use them!