Wavelet U-Net++ for accurate lung nodule segmentation in CT scans: Improving early detection and diagnosis of lung cancer

Abstract

Lung cancer is one of the leading causes of cancer-related deaths globally, and accurate segmentation of lung nodules is critical for its early detection and diagnosis. However, small nodules often have low contrast and are challenging to distinguish from noise and other structures in medical images, making accurate segmentation difficult. In this paper, we propose a new approach called Wavelet U-Net++ for accurately segmenting lung nodules. Our approach combines the U-Net++ architecture with wavelet pooling to capture both high- and low-frequency information in the image, enabling improved segmentation accuracy. Specifically, we use the Haar wavelet transform to downsample the feature maps in the encoder, allowing for fine-grained details in the image to be captured. We evaluated our proposed approach on the LIDC-IDRI dataset, which consists of 1018 CT scans with annotated lung nodules. Our experimental results demonstrate that our approach outperforms several state-of-the-art segmentation methods, achieving a mean dice coefficient of 0.936 and a mean IoU of 0.878. Moreover, we show that wavelet pooling combined with Tversky and CE loss improves the network's ability to detect small and irregular nodules that are conventionally difficult to segment, demonstrating the effectiveness of combining loss functions. Overall, our proposed approach demonstrates the effectiveness of combining wavelet pooling with the U-Net++ architecture for accurate segmentation of lung nodules.