Page 54 - AIH-2-1

Artificial Intelligence in Health Asymmetric U-Net for enhanced spinal MRI segmentation

include details on learning rate, batch size, number of The proven applicability of Dice Loss in medical
epochs, and other hyperparameters critical for training imaging further validates its use in this study. Dice Loss
deep learning models. The segmentation performance is widely employed in medical image segmentation
of J-Unet was benchmarked against several established tasks, particularly those demanding high accuracy, such
models: Unet, Unet++, Unet+++, and Res-Unet. Each as tumor segmentation. Its sensitivity to fine structures
model was trained and evaluated using the same dataset and boundaries makes it an excellent choice for medical
and parameter settings to ensure a fair comparison. The applications, where the accurate segmentation of anatomical
results were analyzed using the aforementioned metrics to structures is critical for diagnosis and treatment planning.
determine the relative strengths and weaknesses of each The combination of the Adam’s optimizer and Dice Loss
model. provides a robust framework for training our segmentation
model. Adam’s efficiency and adaptability, coupled with
3.2. Optimization and loss function selection Dice Loss’s robustness to class imbalance and emphasis on

In this study, we utilized the Adam optimizer due to its accurate segmentation, ensure a high-performance model
significant advantages in deep learning model training. suitable for the complexities of medical image analysis.
Adam’s ability to apply different scaling factors to
parameter updates facilitates the discovery of the global 3.3. Model training
optimum more efficiently during the training process. This During the training process, Dice Loss was employed as
tailored approach to parameter adjustment helps navigate the loss function for our deep learning model. Dice Loss
the complex loss landscapes often encountered in deep is a widely used metric in image segmentation tasks,
learning models, ensuring that each parameter evolves at designed to measure the similarity between the predicted
an appropriate pace. By adaptively adjusting the learning segmentation and the ground truth.
rate for each parameter, Adam accelerates convergence, The segmentation models, namely UNET, UNET++,
enhancing the overall efficiency of the training process. In UNET+++, ResUNET, and J-UNET, were trained using
addition, Adam is known for its computational efficiency the training set. Figure 4 illustrates the variation of loss
and ease of implementation. It requires less memory and Dice coefficient during training and validation for
compared to other optimizers, making it ideal for handling these five models. The graphs provide a clear comparison
large-scale data and complex models. To ensure stable of how each model’s performance evolved over the training
convergence towards a local optimum, we set the learning epochs. During the training process, we encountered
rate to 0.0001. This small learning rate helps in fine-tuning several instances of gradient explosions, particularly with
the model parameters, preventing overshooting and the UNET+++ model. These gradient explosions resulted
ensuring a smooth descent in the loss landscape. in a sharp increase in loss values. To address this issue, we

The choice of Dice Loss as the loss function for this employed a strategy of resuming training from checkpoints
study is driven by its effectiveness in addressing common saved before the occurrence of the gradient explosions.
challenges in medical image segmentation. Dice Loss is This approach allowed us to continue training effectively,
particularly robust in scenarios with imbalanced classes, leading to eventual convergence. The training of the
a common occurrence in medical imaging. Its calculation remaining network models exhibited normal convergence
involves the intersection and union of the predicted and patterns without such disruptions.
true values, making it less sensitive to the disproportionate Upon analyzing the final convergence ranges, it was
pixel counts of different classes. This robustness ensures evident that J-UNET and UNET achieved lower loss
that the model performs well even when certain classes are values on the training set compared to the other models.
underrepresented. On the test set, although the differences in loss values
Moreover, Dice Loss provides smoother gradients among the various models were not substantial, J-UNET
compared to other loss functions, contributing to a more and ResUNET consistently demonstrated smaller loss
stable training process. This stability helps mitigate issues values. This indicates that J-UNET, in particular, exhibited
such as exploding or vanishing gradients, which can hinder superior generalization ability, aligning well with the
the training of deep neural networks. By emphasizing the task objectives. The combination of Dice Loss and the
similarity between predicted and true segmentations, Dice robust architecture of J-UNET contributed to its superior
Loss encourages the model to produce accurate and refined performance in segmenting vertebrae and intervertebral
segmentation results. This focus on overlap and accuracy is discs in spinal MR images. The J-UNET model not only
crucial for tasks requiring precise delineation of structures, converged effectively but also showed a strong ability to
such as in medical imaging where detail and precision are generalize from the training set to the test set, making it a
paramount. highly effective tool for medical image segmentation tasks.

Volume 2 Issue 1 (2025) 48 doi: 10.36922/aih.3889