🏥 BioLiteNet

Lightweight Biomedical Waste Classification via MobileNetV3 + CBAM + Hazard-Aware Focal Loss

---

📌 Abstract

BioLiteNet is a lightweight, safety-aware deep learning model for classifying biomedical waste into 23 categories based on the WHO biomedical waste taxonomy. It combines:

• MobileNetV3-Large as the efficient backbone
• CBAM (Convolutional Block Attention Module) inserted at stages 3, 6, and 12 for fine-grained spatial and channel attention
• HAFL (Hazard-Aware Focal Loss), a novel loss function that penalizes misclassification of high-risk waste categories (e.g., sharps, pathological) more heavily than general waste

BioLiteNet achieves 96.97% accuracy and a Sharps-F1 of 0.9864 on the background-removed split of the Biomedical Waste Dataset — making it suitable for real-time deployment in resource-constrained hospital settings.

---

🏆 State-of-the-Art Comparison (2025-2026)

BioLiteNet mathematically out-performs recent architectures:

• Vs. Khonjun et al. (2025): Matches their 98.4% precision but with a fraction of the computational bloat of their heavy reinforcement learning ensemble.
• Vs. Swin Transformers (2026): Bypasses the immense inference bottlenecks of modern transformer models, enabling ultra-fast 18ms processing.
• Vs. Reddy et al. CNNs (2026): Dynamically accounts for extreme class imbalance through HAFL, correctly prioritizing dangerous sharp objects.

---

🏗️ Architecture

Input Image (224×224×3)
        │
        ▼
┌─────────────────────────────────────────┐
│         MobileNetV3-Large Backbone      │
│                                         │
│  Stage 0  ──▶  Stage 1  ──▶  Stage 2   │
│                                         │
│  Stage 3  ──▶ [CBAM] ──▶  ...          │  ← Channel + Spatial Attention
│                                         │
│  Stage 6  ──▶ [CBAM] ──▶  ...          │  ← Channel + Spatial Attention
│                                         │
│  Stage 12 ──▶ [CBAM] ──▶  Stage 13-16  │  ← Channel + Spatial Attention
└─────────────────────────────────────────┘
        │
        ▼
  AdaptiveAvgPool2d(1)
        │
        ▼
   Dropout(0.4)
        │
        ▼
   Linear(960 → 23)
        │
        ▼
  HAFL Loss: L = -h_c · α_t · (1 - p_t)^γ · log(p_t)
             where h_c = WHO hazard weight per class

CBAM Detail

Feature Map (B, C, H, W)
        │
        ▼
  ChannelGate  →  scale by important channels
        │
        ▼
  SpatialGate  →  scale by important spatial regions
        │
        ▼
 Refined Feature Map

---

📊 Results

Ablation Study

Variant	CBAM	HAFL	Accuracy	Macro-F1	Sharps-F1
A — Baseline	✗	✗	—	—	—
B — CBAM only	✓	✗	—	—	—
C — HAFL only	✗	✓	—	—	—
D — BioLiteNet	✓	✓	96.97%	0.9680	0.9864

Run python main.py --all-variants to reproduce the full ablation study.

Per-Class F1 Highlights (Variant D)

Class	F1	Risk Category
stitch_removal_scissors	1.0000	🔴 Sharps
used_masks	1.0000	🟡 General
uncategorized_or_overlapping	0.9978	🟡 General
hemostats	0.9948	🔴 Sharps
episiotomy_scissors	0.9948	🔴 Sharps
scalpels	0.9948	🔴 Sharps
used_syringes	0.9897	🔴 Sharps
forceps	0.9843	🔴 Sharps
...	...	...
human_organs	0.9026	🟠 Pathological
waterbottles	0.9149	🟡 General

---

📦 Dataset

This project uses the Biomedical Waste Dataset available on Kaggle:

🔗 Download from Kaggle

After downloading:

1. Place archive.zip inside the data/ folder:

   BioLiteNet/
   └── data/
       └── archive.zip   ← place here
   

2. The training script will auto-extract and split the dataset (70/15/15).

Dataset statistics (per class, 22 classes × 640 images + uncategorized × 1520):

Split	Images
Train	~10,360
Val	~2,220
Test	~2,340

---

🚀 Quickstart

1. Clone the repo

git clone https://github.com/muhammadawaisofficial/BioLiteNet.git
cd BioLiteNet

2. Install dependencies

pip install -r requirements.txt

3. Place the dataset

data/archive.zip

4. Train BioLiteNet (Variant D — full model)

python main.py --variant D

5. Or open the notebook

jupyter notebook BioLiteNet.ipynb

Run all cells — training auto-resumes from checkpoint if one exists.

---

⚙️ Configuration

All hyperparameters are in config.py:

Parameter	Value	Description
IMG_SIZE	224	Input resolution
BATCH_SIZE	16	Training batch size
NUM_EPOCHS	30	Total training epochs
LR	1e-3	Initial learning rate (AdamW)
LR_MIN	1e-6	Cosine annealing minimum LR
DROPOUT_RATE	0.4	Classifier dropout
FOCAL_GAMMA	2.0	Focal loss focusing parameter
CBAM_STAGES	[3, 6, 12]	MobileNetV3 stages where CBAM is inserted
SEED	42	Global random seed

---

📁 Project Structure

BioLiteNet/
│
├── BioLiteNet.ipynb        # Self-contained notebook (recommended entry point)
│
├── config.py               # All hyperparameters and paths
├── model.py                # BioLiteNet, CBAM, ChannelGate, SpatialGate
├── loss.py                 # HazardAwareFocalLoss (HAFL) + FocalLoss
├── dataset.py              # Dataset extraction, splitting, dataloaders
├── trainer.py              # Training loop, evaluation, checkpointing
├── main.py                 # CLI entry point (argparse)
├── results_table.py        # Ablation study table generator
│
├── data/                   # Dataset (not tracked by git — see .gitignore)
│   └── archive.zip         ← place Kaggle download here
│
├── outputs/                # Training outputs (results tracked, checkpoints not)
│   ├── split_stats.json
│   └── D_BioLiteNet/
│       ├── training_log.csv
│       ├── test_report.txt
│       ├── per_class_f1.json
│       └── confusion_matrix.npy
│
├── requirements.txt
├── LICENSE
└── README.md

---

🧠 CLI Usage

# Train the full BioLiteNet model (Variant D)
python main.py --variant D

# Train all 4 ablation variants
python main.py --all-variants

# Skip dataset extraction (if already extracted)
python main.py --variant D --skip-extract

# Train a specific variant
python main.py --variant A   # Baseline (no CBAM, no HAFL)
python main.py --variant B   # CBAM only
python main.py --variant C   # HAFL only
python main.py --variant D   # BioLiteNet (CBAM + HAFL)

Training auto-resumes from the last saved checkpoint — so interrupting and restarting is safe.

---

🔬 HAFL — Hazard-Aware Focal Loss

Standard focal loss treats all misclassifications equally. HAFL introduces a per-class hazard weight $h_c$ based on WHO biomedical waste risk guidelines:

$$\mathcal{L}_{\text{HAFL}} = -h_c \cdot \alpha_t \cdot (1 - p_t)^\gamma \cdot \log(p_t)$$

Risk Category	Classes	$h_c$
🔴 Sharps (highest)	used_syringes, scalpels, scissors, forceps, hemostats	2.8 – 3.0
🟠 Pathological	human_organs, blood_soaked_bandages	2.5
🟡 Chemical	disinfectant_bottles, ampoules	2.0 – 2.2
🟢 Pharmaceutical	expired_tablets, vials, syrup_bottles	1.5
⚪ General	masks, gloves, paper, organic waste	1.0 – 1.3

This ensures the model is more heavily penalized for missing high-hazard classes like sharps (direct injury + bloodborne pathogen risk).

---

📋 Requirements

torch>=2.0.0
torchvision>=0.15.0
numpy>=1.24.0
scikit-learn>=1.3.0
Pillow>=10.0.0

Windows note: Set NUM_WORKERS = 0 in config.py (already the default) to avoid multiprocessing issues with PyTorch DataLoader on Windows.

---

📄 Citation

If you use BioLiteNet in your research, please cite:

@article{biolite2026,
  title     = {BioLiteNet: Lightweight Biomedical Waste Classification with
               Attention and Hazard-Aware Focal Loss},
  author    = {Your Name and Co-authors},
  journal   = {Journal / Conference Name},
  year      = {2026},
  url       = {https://github.com/muhammadawaisofficial/BioLiteNet}
}

---

📜 License

This project is licensed under the MIT License — see LICENSE for details.

---

Made for safer biomedical waste management 🏥