Leveraging Modified Cascaded Fully Convolutional Neural Networks and Optimized Swin U-Net for Liver Lesion Prediction

Abstract

Liver lesions are abnormal growths that can develop for various reasons, some of which are noncancerous (benign) while others are cancerous. Machine learning and, deep learning techniques are employed for liver health assessment and lesion detection, but the quality and, diversity of training data can hinder the performance, leading to overfitting or underfitting. These models often need extensive pre-processing and may struggle with accurately segmenting complex patterns. This paper presents an integrated deep learning model for liver lesion prediction that integrates complementary image enhancement, segmentation, optimization and classification techniques. These enhanced images were then segmented using the Swin U-Net (SUN) approach, optimized with the Fossa Optimization Algorithm (FOA) to improve prediction accuracy. The segmented images were subsequently classified using a modified Cascaded Fully Convolutional Neural Network (CFCNN), which incorporated a Conditional Random Field (CRF) in the fully connected layer to ensure high connectivity and reduce computational complexity. Based on the segmented tumor masks, the severity of the disease categorized as low, medium, or high was predicted. The results demonstrate 96.45% accuracy, 96.85% precision, and 96.24% recall. The segmentation achieved an SSIM of 0.96% and a PSNR of 61.97%. Consequently, these methods are well-suited for real-time applications, providing timely and reliable assessments crucial for ensuring the quality of liver lesion detection.