pyMEAL: Multi-Encoder-Augmentation-Aware-Learning

pyMEAL is a multi-encoder framework for augmentation-aware learning that accurately performs CT-to-T1-weighted MRI translation under diverse augmentations. It utilizes four dedicated encoders and three fusion strategies, concatenation (CC), fusion layer (FL), and controller block (BD), to capture augmentation-specific features. MEAL-BD outperforms conventional augmentation methods, achieving SSIM > 0.83 and PSNR > 25 dB in CT-to-T1w translation.

Dependecies

tensorflow

matplotlib

SimpleITK

scipy

antspyx

Available Models

Model ID	File Name	Description
BD	`builder1_mode1l1abW512_1_11211z1p1rt_.h5`	Builder-based architecture model
CC	`best_moderRl_RHID2_1mo.h5`	Encoder-concatenation-based configuration
FL	`bestac22_mode3l_512m2_m21.h5`	Feature-level fusion-based model
NA	`direct7_11ag23f11.h5`	Direct training baseline model
TA	`best_modelaf2ndab7_221ag12g11.h5`	traditional augmentation configuration model

Model Architecture Overview

Figure 1. Model architecture for the model having no augmentation and traditional augmentation.

Figure 2. Model architecture for Multi-Stream with a Builder Controller block method (BD), Fusion layer (FL), and Encoder concatenation (CC).

Download Model Files

You can download any .h5 file directly:

How to Use

Load a Model (Basic)

import tensorflow as tf

# Load the model
model = tf.keras.models.load_model("model.h5", compile=False)

# Run inference
output = model.predict(input_data)

Here, input_data refers to a CT image, and the corresponding T1-weighted (T1w) image is produced as the output.

For detailed instructions on how to use each module of the pyMEAL software, please refer to the tutorial section on our GitHub repository.

How to Get Support

For help, contact:

Dr. Ilyas ([email protected])
Dr. Maradesa ([email protected])