ConvNeXt (AD vs NC) — ADNI Slice Classification

recognition/README.md

@@ -0,0 +1,10 @@
+# Recognition Tasks
+Various recognition tasks solved in deep learning frameworks.
+
+Tasks may include:
+* Image Segmentation
+* Object detection
+* Graph node classification
+* Image super resolution
+* Disease classification
+* Generative modelling with StyleGAN and Stable Diffusion

recognition/adni-convnext-simar/.gitignore

@@ -0,0 +1,34 @@
+# OS
+.DS_Store
+
+# Python
+__pycache__/
+*.pyc
+
+# HPC / SLURM
+*.slurm
+slurm-*.out
+*.out
+*.err
+
+# Training artifacts
+runs/
+checkpoints/
+*.pt
+*.pth
+*.ckpt
+
+# Local notebooks / scratch
+*.ipynb_checkpoints
+
+# Eval byproducts
+recognition/adni-convnext-simar/images/metrics.json
+
+# HPC/SLURM
+*.slurm
+slurm-*.out
+*.out
+*.err
+
+# macOS
+.DS_Store

recognition/adni-convnext-simar/README.md

@@ -0,0 +1,341 @@
+# ConvNeXt for AD vs NC MRI Slice Classification (ADNI)
+
+> version: 1.0    
+> author: Simar Wadhawan    
+> course: COMP3710 – Pattern Analysis & Recognition    
+> repo_url: https://github.com/simarrawadhawan/PatternAnalysis-2025    
+> branch: topic-recognition    
+
+## Summary: 
+This project fine-tunes a ConvNeXt (Small) CNN on a 2-class ADNI slice dataset (AD vs NC). It implements a clean PyTorch training/evaluation pipeline with: label smoothing, weighted loss, 384px inputs, cosine LR schedule with warmup, TTA at test-time
+(orig +hflip), and threshold tuning. Final slice-level test accuracy is ~78% with strong NC recall while improving AD recall over the
+baseline. All code is modularized (dataset.py, train.py, predict.py, modules.py).
+
+### Why Convnext: 
+ConvNeXt is a modern ConvNet that matches or beats transformer baselines (e.g., Swin) when trained on ViT-style design choices (depthwise   convs, GELU, LayerNorm in convnets, larger patch sizes). It remains compute-efficient and stable for grayscale medical imaging.
+
+# Table of Contents:
+  - Introduction
+  - Dataset
+  - Repository Structure
+  - Environment & Setup
+  - Training
+  - Evaluation
+  - Results (slice-level only)
+  - Visualisations
+  - Reproducibility & Determinism
+  - Notes for Markers
+  - References
+  - License
+
+---
+
+# Introduction: 
+  We tackle binary classification of brain MRI slices into Alzheimer’s Disease (AD) vs Normal Control (NC) using a ConvNeXt-S backbone. The   pipeline emphasizes robust training (class weights, label smoothing) and careful evaluation (TTA + calibrated threshold). We report      **slice-level** metrics per the course rubric.
+
+## Dataset:
+  - Name: ADNI (Alzheimer’s Disease Neuroimaging Initiative)
+  - Task: Binary classification on 2D MRI slices: AD vs NC
+  - Classes: ["AD", "NC"]
+  - Approximate_counts:
+    Total Images: 30520
+    Train Images: 21520
+    Test Images: 9000
+  - Balance Note: Roughly balanced per split, mild skew per class.
+
+  ### Expected Directory: 
+
+    ADNI/
+    ├── meta_data_with_label.json
+    └── AD_NC/
+        ├── train/
+        │   ├── AD/    
+        │   └── NC/    
+        └── test/
+            ├── AD/   
+            └── NC/ 
+
+  ### Transforms:
+
+    Train: 
+      - Resize to 384×384
+      - RandomAffine (small degrees/translate/scale)
+      - RandomHorizontalFlip(p=0.5)
+      - ToTensor + Normalize (ImageNet stats for ConvNeXt pretraining)
+    Test: 
+      - Resize to 384×384
+      - ToTensor + Normalize
+    TTA: "Original + horizontal flip, logits averaged"
+
+
+### Repository Structure: 
+  ```text
+  recognition/adni-convnext-simar/
+  ├── dataset.py       # ADNIDataset + transforms
+  ├── modules.py       # build_model() → ConvNeXt variants via timm
+  ├── train.py         # training loop, early-stopping, checkpointing
+  ├── predict.py       # evaluation + visualisations + metrics.json
+  ├── utils.py         # helpers (metrics, samplers, seed utils)
+  ├── requirements.txt # pinned deps
+```
+
+## Environment:
+  - Python: >=3.10  
+  - Requirements_file: requirements.txt  
+  - Key Packages:  
+    - torch, torchvision  
+    - timm  
+    - numpy, pandas  
+    - scikit-learn  
+    - matplotlib, seaborn  
+    - tqdm  
+
+## Installations: 
+    conda create -n comp3710 python=3.10 -y
+    conda activate comp3710
+
+    # install deps
+    pip install -r requirements.txt
+
+ ## Paths:
+  - Data Root: "/home/groups/comp3710/ADNI/AD_NC"
+  - Checkpoint Example: "runs/adni_384_ls01_nomix_last/best.pt"
+  - Image Outputs: "images/"
+
+---
+
+# Training:
+  - Model: ConvNeXt-Small (timm)  
+  - Image Size: 384  
+  - Batch Size: 32  
+  - Epochs: 30-60  
+  - Optimizer: AdamW  
+  - Scheduler: CosineAnnealingLR with warmup  
+  - Learning Rate: 5.0e-5  
+  - Weight Decay: 0.05  
+  - Label Smoothing: 0.1  
+  - Class Weights: "Computed from train split (AD vs NC)"  
+  - Mixup: Disabled in final 78% run  
+  - Early Stopping:  
+     - Monitor: "val_acc"  
+     - Patience: 10  
+  - Freeze Policy: "Unfreeze last 3 stages + head (≈99.5% trainable here)"  
+
+  ## Commands: 
+    python train.py \
+      --data-root /home/groups/comp3710/ADNI/AD_NC \
+      --model convnext_small \
+      --img-size 384 \
+      --batch-size 32 \
+      --epochs 60 \
+      --lr 5e-5 \
+      --weight-decay 0.05 \
+      --label-smoothing 0.1 \
+      --no-mixup \
+      --save-dir runs/adni_384_ls01_nomix_last
+
+---
+
+# Evaluation:
+ - Mode: Slice-level (primary, per rubric)  
+ - TTA: orig + hflip (averaged)  
+ - Threshold: P(NC) = 0.55 used for the 78.18% run  
+
+ # Outputs: 
+
+    images/
+    ├── confusion_matrix.png
+    ├── roc_curve.png
+    ├── performance_metrics.png
+    ├── sample_predictions.png
+    ├── misclassified_samples.png
+    └── metrics.json
+
+  ## Commands: 
+    python predict.py \
+      --checkpoint runs/adni_384_ls01_nomix_last/best.pt \
+      --data-root /home/groups/comp3710/ADNI/AD_NC \
+      --model convnext_small \
+      --batch-size 32 \
+      --img-size 384 \
+      --save-dir images \
+      --num-workers 4
+
+  - Download from Rangpur: 
+    > scp -r [email protected]:/home/Student/s4977354/PatternAnalysis-2025/recognition/adni-convnext-simar/images ./
+
+---
+
+# Results Slice Level:
+ ## Overall:
+  - Accuracy: 0.7818
+  - precision_weighted: 0.7923
+  - recall_weighted: 0.7818
+  - f1_weighted: 0.7795
+  - specificity_AD: 0.6827
+  - auc_roc: 0.8497
+  - total_samples: 9000
+
+ ## Per Class:
+  - AD:
+    - Precision: 0.8472
+    - Recall: 0.6827
+    - f1: 0.7561
+    - Support: 4460
+
+  - NC:
+    - Precision: 0.7383
+    - Recall: 0.8791
+    - f1: 0.8025
+    - Support: 4540
+
+  - Confusion Matrix Counts:
+    - TN_AD_correct: 3045
+    - FP_AD_as_NC: 1415
+    - FN_NC_as_AD: 549
+    - TP_NC_correct: 3991
+
+  Note: 
+  > Metrics above correspond to TTA (orig+hflip) with threshold P(NC)=0.55.
+    This configuration yielded the best slice-level balance in our runs.
+
+---
+
+# Figures:
+
+   ## Results — Visualizations (Slice-Level, Threshold = 0.55, TTA = orig+hflip)
+
+<p align="center">
+  <img src="images/confusion_matrix.png" alt="Confusion Matrix (Slice-Level)" width="64%">
+  <br/>
+  <sub><b>Figure 1.</b> Confusion Matrix — slice-level.</sub>
+</p>
+
+**What it shows (slice-level):**  
+Accuracy **78.18%** (7036/9000). True AD (TN)=**3045**, False AD→NC (FP)=**1415**; True NC (TP)=**3991**, False NC→AD (FN)=**549**. Model is slightly conservative for NC (higher NC recall), with most errors being AD→NC.
+
+---
+
+<p align="center">
+  <img src="images/roc_curve.png" alt="ROC Curve (Slice-Level)" width="64%">
+  <br/>
+  <sub><b>Figure 2.</b> ROC Curve — slice-level.</sub>
+</p>
+
+**What it shows:**  
+Overall discrimination between AD and NC. The curve sits well above the diagonal (random). AUC summarizes ranking performance; even when the threshold is fixed at 0.55 for reporting, the ROC shows performance across all thresholds.
+
+---
+
+<p align="center">
+  <img src="images/performance_metrics.png" alt="Performance Metrics (Slice-Level)" width="60%">
+  <br/>
+  <sub><b>Figure 3.</b> Metric Summary — slice-level.</sub>
+</p>
+
+**What it shows:**  
+Accuracy **0.7818**, Precision **0.7923**, Recall **0.7818**, F1 **0.7795**, Specificity **0.6827**. Balanced Precision/Recall, with Specificity lower than Sensitivity → the model favors catching NC over rejecting AD false alarms.
+
+---
+
+<p align="center">
+  <img src="images/sample_predictions.png" alt="Sample Predictions (Correct & Incorrect)" width="90%">
+  <br/>
+  <sub><b>Figure 4.</b> Sample Predictions — slice-level.</sub>
+</p>
+
+**What it shows:**  
+A random subset of predictions with model confidence. Green titles are correct; red would indicate errors. This provides qualitative insight into what the model considers “confidently” AD vs NC at the slice level.
+
+---
+
+<p align="center">
+  <img src="images/misclassified_samples.png" alt="Misclassified Samples (Slice-Level)" width="90%">
+  <br/>
+  <sub><b>Figure 5.</b> Misclassified Samples — slice-level.</sub>
+</p>
+
+**What it shows:**  
+Representative failures (always red titles). Typical mistakes include anatomically subtle slices (early AD or near-blank edge slices) where class cues are weak. These errors motivate aggregation or weighting by slice informativeness in future work.
+
+---
+
+<p align="center">
+  <img src="images/confusion_matrix_roc.png" alt="Combined Confusion Matrix + ROC (Slice-Level)" width="90%">
+  <br/>
+  <sub><b>Figure 6.</b> Combined CM + ROC — slice-level (for compact reporting).</sub>
+</p>
+
+**What it shows:**  
+A single panel to include in reports: confusion matrix counts plus ROC curve. Useful when space is limited but both error types and discrimination performance must be shown.
+
+---
+
+### Metrics (Slice-Level, for Repro)
+- **Accuracy:** 78.18%  
+- **Precision (weighted):** 0.7923  
+- **Recall (weighted):** 0.7818  
+- **F1 (weighted):** 0.7795  
+- **Specificity (AD as negative class):** 0.6827  
+- **Counts:** TN=3045, FP=1415, FN=549, TP=3991  
+- **Setup:** ConvNeXt Small @ 384px, label smoothing=0.1, **no mixup**, TTA (orig + hflip), threshold=0.55
+
+---
+
+# Usage Quickstart: 
+
+  ### 1) Install
+  conda activate comp3710
+  pip install -r requirements.txt
+
+  ### 2) Train
+  python train.py --data-root /home/groups/comp3710/ADNI/AD_NC --img-size 384 --batch-size 32
+
+  ### 3) Evaluate (slice-level + figures)
+  python predict.py --checkpoint runs/adni_384_ls01_nomix_last/best.pt --save-dir images
+
+  ### 4) Copy figures off Rangpur
+  scp -r [email protected]:/home/Student/s4977354/PatternAnalysis-2025/recognition/adni-convnext-simar/images ./
+
+## Checklist To Submit: 
+  - [x] Push updated .py files (no checkpoints)
+  - [x] README with slice-level 78% results and figures
+  - [x] Open PR from topic-recognition → main with clear title/body
+  - [x] Export README to PDF (GitHub “Print to PDF” or Markdown → PDF)
+  - [x] Submit PDF + repo link
+
+---
+
+# References:
+  - Name: ConvNeXt  
+    Link: "https://arxiv.org/abs/2201.03545"
+  - Name: timm: PyTorch Image Models  
+    Link: "https://github.com/huggingface/pytorch-image-models"
+  - Name: ADNI overview  
+    Link: "https://adni.loni.usc.edu/"
+  - Name: ROC/AUC interpretation (medical)  
+    Link: "https://pmc.ncbi.nlm.nih.gov/articles/PMC12260203/"
+  - ### AI Assistance & Authorship
+I used **ChatGPT (GPT-5 Thinking)** as a writing/coding assistant for this project.  
+**Scope of assistance:**
+- Helped draft and polish this README (structure, wording, and formatting).
+- Suggested code refactors and guardrails (e.g., dataloader fixes, evaluation logging, commit message style).
+- Generated shell/Git one-liners and SLURM templates, which I **reviewed and edited**.
+- Provided troubleshooting ideas (e.g., threshold sweep, TTA, majority-vote scan aggregation) which I **implemented and verified**.
+
+**Not done by AI:**
+- Dataset preparation, training runs, hyperparameter choices, model selection, and **all results** were executed by me.  
+- Figures (confusion matrix, ROC, metrics bar chart, samples) were generated directly from my evaluation scripts and saved under `recognition/adni-convnext-simar/images/`.
+
+**Verification & integrity:**
+- I reviewed every AI suggestion before use and tested all code paths that entered the repository.
+- No metrics or figures were fabricated or manually edited; they are reproducible from the provided scripts and logs.
+- External sources are cited in the **References** section.
+
+**Provenance (dates/tools):**
+- ChatGPT sessions: Oct–Nov 2025  
+- Model: GPT-5 Thinking (ChatGPT)  
+- Purpose: drafting + developer productivity; **final technical decisions are mine**.
+
+
+License: Apache-2.0 (same as course starter)