Abstract:
This paper proposes an automated nasal sinus disease segmentation system
using an enhanced U-Net model with VGG16 backbone and attention
mechanism. It would overcome the weaknesses inherent in manual
segmentation by presenting a superior and accurate solution. The data set for
this research consists of medical image data, public and self-collected,
preprocessed for segmentation purposes. Preprocessing pipeline included data
annotation using Roboflow, resizing, normalization, and binary mask creation.
Random rotation, flipping, and shifting were some of the data augmentation
methods used to make the model more robust. The proposed model architecture
includes a VGG16-based encoder with a bottleneck for enhancing features and a
decoder with attention gates to achieve maximum attention in regions of interest
in disease areas. Adam optimizer and binary cross-entropy loss were used for
training the model and tested on Dice coefficient, Intersection over Union (IoU),
and F1-score. The model underwent great improvement after 50 epochs with
accuracy 98.43%, which indicates its proficiency in nasal sinus disease
segmentation. Furthermore, the proposed model was compared with other
existing segmentation models, and the results demonstrated its superior
performance, justifying the effectiveness of the custom architecture.
