MammoFusionNet: A unified approach for breast lesion classification and segmentation using modified involutional neural networks

Abstract

Breast cancer impacts more than 1.5 million people worldwide annually, with early detection essential for differentiating between malignant and benign lesions. The precision of diagnostics frequently relies on the proficiency of healthcare practitioners, resulting in possible misdiagnoses. Consequently, the development of efficient computer-aided procedures is imperative for enhancing early detection and decreasing death rates. Although numerous research concentrate on lesion identification from ultrasound images, they often tackle classification or segmentation independently. To address this issue, we present MammoFusionNet, a multi-task learning architecture that simultaneously executes the classification and segmentation of breast lesions. MammoFusionNet integrates a modified Involution Neural Network (INN) into a UNet architecture, enhancing its flexibility to many imaging modalities and resolutions. To validate the model, we used the Breast Ultrasound Images Dataset (BUSI) which consists of two classes: Benign and Lesion. MammoFusionNet exceeds the performance of leading models such as InceptionResNetV2, EfficientNetB7, and ResNet50 in classification tasks, attaining 99.12% accuracy, 98% AUC, and a 99% F1-score. Additionally, it surpasses segmentation models like Swin-Unet and DeepLabV3, achieving an 82% IOU, 90% mean IOU, and a 75% Dice score. The integration of classification and segmentation tasks improves overall performance, establishing MammoFusionNet as an effective instrument for breast cancer diagnosis in clinical environments. © 2025 University of Bahrain. All rights reserved.