Temporal and Spatial Fusion for Breast Lesion Classification in Ultrasound Videos using Hybrid UNet-LSTM Architecture

This research involves the design, implementation, and evaluation of a Hybrid UNet-LSTM architecture aimed at enhancing the classification of breast lesions in ultrasound videos, which is critical for the early and accurate diagnosis of breast cancer. The model integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks inspired from the U-Net architecture to capture both spatial and temporal features from video sequences. This approach addresses the inherent challenges of ultrasound imaging, such as speckle noise, operator dependency, and variability in lesion appearances, which often complicate diagnosis. The Hybrid UNet-LSTM model processes ultrasound videos, employing intra-video and inter-video fusion techniques to improve classification accuracy. The model is built using the latest advancements in deep learning, ensuring a robust and reliable system for medical diagnostics. The implementation of this Hybrid UNet-LSTM architecture serves as a foundational model that can be further improved upon to develop more advanced models in the future, enhancing breast cancer diagnostics by providing a more accurate and reliable method for lesion classification. This groundwork supports ongoing innovation in the field, with the potential to improve patient outcomes and assist radiologists in clinical decision-making.