medscan-cu124 0.1.1

Medscan with CUDA 12.4 (PyTorch GPU support)

MedScan – End‑to‑End Medical Imaging Training Pipeline

MedScan is a self‑contained Python package that lets you pre‑process and augment data, train deep‑learning models, evaluate with plots/metrics, and save / reload checkpoints – all from one intuitive API or a single command‑line call.

---

Table of Contents

1. Key Features
2. Repository Layout
3. Installation
4. Quick‑Start Notebook
5. Command‑Line Usage
6. Prediction CLI
7. API Overview
8. Training Pipeline Walk‑Through
9. Saving, Loading & Plot Outputs
10. Troubleshooting
11. License

---

Key Features

• One‑line data split with Data.split() (stratified, group‑aware, or plain).

• Config‑driven augmentation, balancing, masking & context‐feature handling via PreprocessConfig.

• Flexible training:

  • Multi‑head model (shared backbone, one head per target) or
  • Single‑head‑per‑label models (optionally across multiple backbones).

• Torch AMP support (mixed precision) & automatic GPU/CPU selection.

• Automatic early stopping per head or per model.

• Rich evaluation with confusion matrices, loss & LR curves, AUC/Accuracy/F1.

• Full CLI runner (train.py) – reproduce your notebook runs head‑less.

• Saved plots are dropped into a plots/ folder automatically when using the CLI.

---

Repository Layout

medscan/                  # Package root
├── __init__.py           # Exposes Data, PreprocessConfig, TrainConfig, Pipeline
├── config.py             # @dataclass configs used throughout
├── data.py               # Data.split utility (train/val/test)
├── pipeline.py           # Core training/inference/evaluation pipeline
└── transform.py          # Augmentation & class‑balancing logic
examples/                 # Example notebooks + sample CSV
  └── merged_dataframe.csv
train.py                  # CLI interface wrapping the Pipeline
README.md                 # (this file)

---

Installation

Prerequisites: Python ≥ 3.9

Target	Command
CPU (default)	pip install medscan
cu116 build	pip install medscan-cu116
cu117 build	pip install medscan-cu117
cu118 build	pip install medscan-cu118
cu121 build	pip install medscan-cu121
cu124 build	pip install medscan-cu124

Pick exactly one line that matches your CUDA toolkit (or none for CPU). No extra wheels needed — each tag bundles the correct Torch + torchvision wheels.

Environment	Requirements file
CPU (default)	pip install -r requirements_cpu.txt
cu116	pip install -r requirements_cu116.txt
cu117	pip install -r requirements_cu117.txt
cu118	pip install -r requirements_cu118.txt
cu121	pip install -r requirements_cu121.txt
cu124	pip install -r requirements_cu124.txt

Tip – if you already have PyTorch installed, make sure the wheel versions match the list above.

---

Quick‑Start Notebook

Open examples/medscan_quickstart.ipynb and follow the annotated steps — the only manual preparation is to create a single DataFrame that already contains img_path plus all target columns. Everything afterwards (split, augment, train, evaluate) is automated by the pipeline.

import pandas as pd
from medscan import Pipeline, Data, PreprocessConfig, TrainConfig

# 1️⃣  Prepare your own merged DataFrame -> df_merged (img_path + labels)
# 2️⃣  Split, configure, train, evaluate — see the notebook for details
# ... see the full notebook in /examples for detailed comments

The notebook in examples shows how to build a toy df_merged, configure preprocessing/training, train for one epoch, evaluate, and save the best model.

---

Command‑Line Usage

train.py wraps every step so you can train from the shell – no notebook needed.

Minimal run

python train.py \
  --data_path "path/to/merged_dataframe.csv"

This uses all defaults: CPU, 70 / 15 / 15 train‑val‑test split, no augmentation, single resnet34 backbone, multi‑head mode, 10 epochs, and saves plots to ./plots/.

Full run (every flag)

python train.py \
  --data_path "path/to/merged_dataframe.csv" \
  --filter_column "Neuro_Imaging=1" \
  --filter_column "Hemisphere=0" \
  # (repeat --filter_column to apply multiple conditions)
  --targets "Neuro_Imaging,Motion_Artefact,Skull_Visibility,Projection,Contrast_fluid,DSA,Hemisphere,ICA_Top_visible,MCA_visible" \
  --train_frac 0.7  \
  --val_frac 0.15  \
  --seed 123 \
  --augment  \
  --augment_factor 2  \
  --balance_on \
  --augmented_image_path augmented_images \
  --elastic_alpha 34.0  \
  --elastic_sigma 4.0 \
  --contrast_min 0.4 \
  --contrast_max 0.9 \
  --input_size 224 \
  --batch_size 32 \
  --epochs 10 \
  --early_stopping_patience 3 \
  --learning_rate 0.001 \
  --optimizer AdamW \
  --dropout \
  --dropout_rate 0.5 \
  --mixed_precision \
  --save_best_model \
  --checkpoint_dir checkpoints \
  --metric val_loss \
  --metric_mode min \
  --confidence_score \
  --pretrained_models "resnet34,resnet50" \
  --train_per_label \
  --force_cpu \
  --save_model_path best_model.pt \
  --plots "confusion_matrix,loss_vs_epoch,lr_vs_epoch" \
  --metrics "AUC,accuracy,F1"

All arguments are documented via python train.py -h.

Plots are saved to ./plots/plot_*.png (auto‑created).

Augmented images are saved under the folder you specify in --augmented_image_path (default: augmented_images/).

---

Prediction CLI

predict.py lets you run inference on a folder of images using a saved pipeline directory or checkpoint.

python medscan/predict.py \
  --img_dir path/to/image_folder \
  --model_path path/to/pipeline_dir \
  --labels all \
  --output_csv predictions.csv \
  --force_cpu

Use --labels to restrict targets, --confidence to output probability scores, and omit --force_cpu to use a GPU if available.

---

API Overview

Class / Function	Role
Data.split	Stratified (or group‑aware) train/val/test split in one call.
PreprocessConfig	Declarative augmentation & balance settings.
TrainConfig	Training hyper‑parameters, device, backbone list, etc.
Pipeline	Orchestrates training, prediction, evaluation, save/load.
transform.augment_and_balance	Internal helper to expand/upsample data.

All objects live under medscan and are re‑exported via __init__.py for convenience:

from medscan import Data, PreprocessConfig, TrainConfig, Pipeline

---

Training Pipeline Walk‑Through

1. Provide a merged DataFrame — user‑supplied, must include img_path and all label columns.

2. Split into train / val / test with Data.split() (ensures each class appears in every subset).

3. Preprocess (PreprocessConfig)

   • Optional augmentation: elastic, contrast, contrast + elastic.
   • Optional class balancing (upsampling) – requires augment=True.
   • Optional mask handling & context features.

4. Model training (Pipeline.fit)

   • Multi‑head (default): one backbone + multiple classification heads.
   • Per‑label: one backbone per target (single head each).
   • Early stopping is tracked individually per head.

5. Prediction (Pipeline.predict): adds Label_<target> columns.

6. Evaluation (Pipeline.evaluate): calculates metrics & shows / saves plots.

7. Save / Load (Pipeline.save / Pipeline.load): preserves all weights and head mapping.

---

Saving, Loading & Plot Outputs

• Saving: model.save("best_model.pt") stores either a single multi‑head state or a dict of per‑label states.
• Loading: supply the same PreprocessConfig / TrainConfig (device can differ) and call model.load().
• Plots: when run via the CLI, every plt.show() call is monkey‑patched to dump PNGs under plots/. In notebooks they still display inline.

---

Troubleshooting

Issue	Fix
CUDA library not found	Install a matching requirements_cuXXX.txt wheel or pass --force_cpu.
Class missing in test/val	Lower train_frac / val_frac or disable require_all_classes.
No plots on server	Use the CLI; plots are saved to disk instead of shown.
OOM on GPU	Reduce --batch_size, --input_size, or train on CPU.

---

License

This project is released under the MIT License. See LICENSE for details.