U-Net for Lung Segmentation (Model 3B – Hard Example Fine-Tuning)

This model performs lung segmentation on grayscale chest X-ray images using a custom deep learning architecture based on U-Net.
It integrates modern components like ASPP, Squeeze-and-Excitation (SE) blocks, and dilated residual convolutions to improve accuracy and generalization in complex medical imaging tasks.

This version (Model 3B) was fine-tuned specifically on hard examples, i.e., structurally difficult or degraded images with known weaknesses from earlier models.

🔧 Architecture Overview

Base Architecture: Extended U-Net
Encoder: 6-level hierarchical encoder with:

Dilated residual blocks (multi-scale context, ResNet-style)
SE blocks (channel-wise attention)
MaxPooling2D at each stage

Bridge: ASPP module with dilation rates of 6, 12, 18 + Global Average Pooling

Decoder:

Upsampling layers with skip connections
Re-applied dilated residual blocks
Progressive channel reduction (1024 → 32)

Output Layer:

Final 1×1 Conv2D + Sigmoid
Binary mask with pixel values in [0, 1]

🧪 Training Configuration

Component	Setting
Optimizer	Adam (lr = 1e-4)
Loss Function	Combined Dice + Binary Crossentropy
Scheduler	ReduceLROnPlateau (patience = 3)
Early Stopping	Monitored on `val_loss`
Epochs (initial phase)	20
CLAHE Preprocessing	Yes (local contrast enhancement)

🎯 Hard Example Fine-Tuning (Model 3B)

After baseline training, this model was further fine-tuned on a subset of 300 challenging samples, automatically selected based on their lowest Dice scores out of 21,165 predictions. These cases included:

Heavily rotated or skewed images
Very noisy or low-contrast scans
Fragmented lung regions and anatomical irregularities

Only model weights were updated – no changes were made to the architecture. Training was resumed using the saved model_3.h5.

📊 Results (Epoch 50)

Metric	Value
Train Dice	0.9818
Train IoU	0.9643
Validation Dice	0.9594
Validation IoU	0.9226
Combined Loss	0.0590
Learning Rate (final)	2.5e-06

✅ Performance Summary

Model demonstrates strong segmentation accuracy on complex or degraded images.
CLAHE preprocessing significantly improved edge detection and contrast in ambiguous regions.
Evaluation across full test set yielded:
- Avg Dice Score: 0.9768
- Avg IoU Score: 0.9551
- Avg Combined Loss: 0.0649

🔍 Usage

from huggingface_hub import hf_hub_download
import tensorflow as tf

# Download model file
model_path = hf_hub_download(repo_id="username/model-3b-lung-segmentation", filename="entwickeltes_masken_model.h5")

# Load model
model = tf.keras.models.load_model(model_path, compile=False)
model.compile(...)  # Add loss/metrics if needed