This model is a fine-tuned version of a custom U-Net architecture enhanced with ASPP(Atrous Spatial Pyramid Pooling), Squeeze-and-Excitation blocks, and dilated convolutions.

It is designed for semantic segmentation of lung regions in chest X-ray images.

The model was further trained on a curated set of difficult X-ray examples with low contrast, overlapping anatomical structures, or weak/incomplete ground truth annotations.

To improve structural realism, the model was fine-tuned using a PatchGAN discriminator in an adversarial training setup.

This encouraged the generation of sharper, more anatomically consistent masks, especially on noisy or edge-case images.
The model also outperforms the original ground-truth masks on visual quality and edge precision.

It was trained on the publicly available COVID-19 Radiography Database.

The final model achieves a Dice score of 95.9% on the internal validation set.

Model Format & Loading Instructions

This model contains custom loss functions and metrics, so you must load it using custom_objects in Keras

Example: Load the model (Keras 3+)

def dice_coefficient(y_true, y_pred): numerator = 2 * K.sum(y_true * y_pred) + 1e-6 denominator = K.sum(y_true + y_pred) + 1e-6 return numerator / denominator

def iou_metric(y_true, y_pred): intersection = K.sum(y_true * y_pred) union = K.sum(y_true) + K.sum(y_pred) - intersection return (intersection + 1e-6) / (union + 1e-6)

def dice_loss(y_true, y_pred): return 1 - dice_coefficient(y_true, y_pred)

def combined_loss(y_true, y_pred): return 0.5 * dice_loss(y_true, y_pred) + 0.5 * tf.keras.losses.binary_crossentropy( y_true, y_pred )

custom_objects = { "dice_coefficient": dice_coefficient, "iou_metric": iou_metric, "dice_loss": dice_loss, "combined_loss": combined_loss, }

model = keras.models.load_model( "model.keras", custom_objects=custom_objects, )