AttentionFractalCovNet Architecture for Breast Ultrasound Image Segmentation

• Published in: 2023 OITS International Conference on Information Technology (OCIT) pp. 94 - 101
• Main Authors: Agarwalla, Rhythm, Ray, Satyajeet Kumar, Paul, Saurav
• Format: Conference Proceeding
• Language: English
• Published: IEEE 13-12-2023
• Subjects: convolution; Fractals; Image segmentation; Lesions; malignant; segmentation; Testing; Tumor; Ultrasonic imaging; Ultrasonic variables measurement; ultrasound; X-ray imaging
• DOI: 10.1109/OCIT59427.2023.10430724

Tumor in the breast is a significant cause of unhealthiness among women. It is primarily the malignant-natured lesions that are a big reason for death via cancer among females. Segmentation of the tumor is complex due to the hazy boundaries and non-identical shapes and sizes. In this paper, an attention-based architecture, called AttentionFractalCovNet(AttFCN) architecture is proposed for the accurate and automatic segmentation of tumors. It takes inspiration from FractalCovNet architecture that forms a W-shape rather than the U-shape of the UNet architecture. Convolution blocks along with Attention Gates(AG) in the second decoding layer are used for achieving better segmentation results. Breast Ultrasound (BUS) images are used as input to segment out the lesions. Various matrices like Accuracy, MSE, Dice Coefficient(DC), Recall, Intersection over Union(IoU), Precision, and F1 Score(F1) were calculated for all the implemented architectures. It achieves higher DC and IoU values as compared to other architectures like SegNet, UNet, Re-sUNet, AttentionUNet(AU), and FractalCovNet(FracCN). Thus, the model is more suitable for the segmentation of tumors and can give better results than the existing architectures mentioned above.