Sidharth1743/Fingernail_anemia_QLora

Fingernail MedSigLIP QLoRA

Overview

This repository provides QLoRA fine‑tuning artifacts for fingernail‑based anemia detection using a MedSigLIP vision encoder. It includes the trained LoRA adapters, linear head, processor config, and detailed evaluation metrics. The goal is a lightweight, GPU‑efficient fine‑tuning pipeline suitable for limited VRAM setups.

Contents

• fingernails_qlora/
  • lora_adapters/ – LoRA adapter weights and config
  • linear_head.pt – classification head
  • processor/ – preprocessing config
  • test_metrics.json – evaluation metrics
  • train_config.json – training hyperparameters
• code/fingernails/ – training and split scripts (mirrors the code repo)
• README.md – technical report

Dataset (Counts Only)

Source dataset folder in the main project:

• Dataset/Detection of Anemia using Colour of the Fingernails Image Datasets from Ghana/Fingernails/

Split counts (from code/fingernails/*.csv):

• Train: 3409 images (Non‑Anemia 1371, Anemia 2038)
• Val: 390 images (Non‑Anemia 150, Anemia 240)
• Test: 417 images (Non‑Anemia 174, Anemia 243)

Method

• Backbone: MedSigLIP vision encoder (frozen backbone).
• Fine‑tuning: QLoRA (LoRA adapters for parameter‑efficient tuning).
• Classifier: linear head on pooled vision embeddings.
• Preprocessing: resize to 448x448, RGB conversion, normalize to [-1, 1].

Experiments & Results

QLoRA Fine‑Tuning (Test Metrics)

From fingernails_qlora/test_metrics.json:

• Accuracy: 0.582734
• Precision: 0.582734
• Recall: 1.000000
• F1: 0.736364
• ROC‑AUC: 0.620288

Zero‑Shot Baselines (Comparison)

Baseline (fingernails, test=417)

• Accuracy: 0.417266
• Precision: 0.000000
• Recall: 0.000000
• F1: 0.000000
• ROC‑AUC: 0.440542
• Confusion Matrix: [[174, 0], [243, 0]]

Baseline (ROI, test=417)

• Accuracy: 0.412470
• Precision: 0.478261
• Recall: 0.090535
• F1: 0.152249
• ROC‑AUC: 0.406485
• Confusion Matrix: [[150, 24], [221, 22]]

Baseline (all fingernails, test=4216)

• Accuracy: 0.439042
• Precision: 0.701031
• Recall: 0.107894
• F1: 0.187006
• ROC‑AUC: 0.564913
• Confusion Matrix: [[1579, 116], [2249, 272]]

Comparison Table (Test Metrics)

Experiment	Accuracy	Precision	Recall	F1	ROC‑AUC	Notes
Zero‑shot baseline (test=417)	0.417266	0.000000	0.000000	0.000000	0.440542	no training
Zero‑shot baseline (ROI, test=417)	0.412470	0.478261	0.090535	0.152249	0.406485	ROI cropped
Zero‑shot baseline (all, test=4216)	0.439042	0.701031	0.107894	0.187006	0.564913	all images
QLoRA fine‑tune (test=417)	0.582734	0.582734	1.000000	0.736364	0.620288	5 epochs

Training Command (QLoRA)

uv run python src/fingernails/qlora_finetune.py \
  --train-csv "src/fingernails/train.csv" \
  --val-csv "src/fingernails/val.csv" \
  --test-csv "src/fingernails/test.csv" \
  --model-dir medsiglip \
  --output-dir results/fingernails_qlora \
  --batch-size 2 \
  --accumulation-steps 8 \
  --epochs 5 \
  --lr 2e-4 \
  --class-weight 1.5 \
  --log-every 20

Hardware Requirements

Minimum working configuration used in this project:

• GPU: 1x NVIDIA RTX 3060 (12 GB VRAM)
• RAM: 16 GB or more recommended
• Disk: 5+ GB (dataset + artifacts)

Limitations

• For research and triage only; not for clinical diagnosis.
• Metrics depend on dataset distribution and acquisition conditions.
• QLoRA recall is high in this run; calibration should be validated in broader settings.

Contact

Model author: Sidharth (Hugging Face: Sidharth1743).