ConvNeXt (AD vs NC) — ADNI Slice Classification

[simarrwadhawan]

Author: Simar Wadhawan • Course: COMP3710 • Branch: topic-recognition

Summary

This PR adds a complete ConvNeXt-Small pipeline to classify ADNI MRI slices into AD vs NC.
Includes dataloader, training, TTA evaluation, and auto-generated plots.
Slice-level test accuracy: 78.18% (with TTA + threshold 0.55).

What’s in this PR

• dataset.py – AD/NC slice dataset + transforms
• train.py – training (AdamW, cosine LR, label smoothing, class weights)
• predict.py – evaluation with TTA + threshold tuning, saves metrics/plots
• modules.py – ConvNeXt build + selective unfreeze
• utils.py – helpers
• images/ – generated figures (confusion matrix, ROC, samples, metrics)

Environment

• Python 3.10+, PyTorch 2.x, CUDA 11.8+
• pip install -r recognition/adni-convnext-simar/requirements.txt

Data (ADNI)

• Not included. Requires ADNI access.
• Expected layout: /home/groups/comp3710/ADNI/AD_NC/{train,test}/{AD,NC}/*.jpeg

Reproducibility

• Seeds fixed in train.py (torch, numpy, python) — results vary ±1%.
• Single-GPU A100 40GB; ~25–40 min for 20 epochs (your mileage may vary).

Quick Run

# Train
python recognition/adni-convnext-simar/train.py --img_size 384 --batch_size 32

# Evaluate (slice-level)
python recognition/adni-convnext-simar/predict.py \
  --checkpoint runs/adni_384_ls01_nomix_last/best.pt \
  --tta --threshold 0.55 \
  --save_dir recognition/adni-convnext-simar/images

Reviewer quick check

# 1) Install
pip install -r recognition/adni-convnext-simar/requirements.txt

# 2) Evaluate with my checkpoint path (adjust if different)
python recognition/adni-convnext-simar/predict.py \
  --checkpoint runs/adni_384_ls01_nomix_last/best.pt \
  --tta --threshold 0.55 \
  --save_dir recognition/adni-convnext-simar/images
  
  

[yexincheng]

This is an initial inspection, no action is required at this point

Recognition Problem : total : 11

1. Solves problem: The solution is appropriate for the problem, achieving an accuracy of 78.18% on the test set.
   The dataset was properly split.
   No loss and metrics plots for training set, but reasonable metrics plots for testing set. (2/5)
2. Implementation functions: Good (3/3)
3. Good design: Well-designed (1/1)
4. Commenting: Insufficient and minimal comments; should include input and output. (0/1)
5. Difficulty: Hard, but imported pre-defined model architecture and used pre-trained checkpoint (5/10)

Note:

• Nice and detailed pull request message.
• Some commits happened after the deadline. (Friday, 31 Oct)
• Almost no model description.
• Nice comparison between slice-level and subject-level.

[gayanku]

Marking

Good/OK/Fair Practice (Design/Commenting, TF/Torch Usage)
	No design and implementation.	-2
	Spacing and comments.
	No Header blocks.	-1
Recognition Problem
	OK solution to problem.	-3
	Driver Script present.
	File structure present.
	Good Usage & Demo & Visualisation & Data usage.
	Module present.
	Commenting present.
	No Data leakage found.
	Difficulty : Normal. Hard. ConvNeXt. Imported pre-trained	-5
Commit Log
	Good Meaningful commit messages.
	Good Progressive commits.
Documentation
	Readme :Acceptable.	-2
	Model/technical explanation :Good.
	Description and Comments :Good.
	Markdown used and PDF submitted.
Pull Request
	Successful Pull Request (Working Algorithm Delivered on Time in Correct Branch).
	No Feedback required.
	Request Description is good.
TOTAL		-13

Marked as per the due date and changes after which aren't necessarily allowed to contribute to grade for fairness.
Subject to approval from Shakes

[wangzhaomxy]

Simar Wadhawan s4977354