ConvMixer-based encoder and classification-based decoder architecture for breast lesion segmentation in ultrasound images

Abstract

Automatic breast lesion segmentation in ultrasound images is an important research topic, as breast cancer is one of the most common and dangerous cancers. However, lesion segmentation is a difficult task due to the challenges encountered in ultrasound images. In this study, a new encoder-decoder network based on ConvMixer is designed for breast lesion segmentation in ultrasound images. This model, called the ConvMixer-based Encoder-Classification-Based Decoder (CE-CD), divides the pixel-level segmentation task into image-level classification and pixel-level detection, effectively combining them. ConvMixer and DenseNet121 are used in the encoder. While spatial and semantic details are obtained with DenseNet121, long-range-context details are obtained with ConvMixer. Then, these features are combined and transferred to the decoder. In addition, the decoder consists of a classification network and a detection network. The detection network obtains the lesion detection score at the pixel level, while the classification network obtains the lesion classification score at the image level. In the last section of CE-CD, the detected lesion class is determined using the classification output with the result generation algorithm. The BUSI dataset was used to analyze the performance of the CE-CD. As a result of experimental studies, the proposed model provided a superior performance than the state-of-the-art models with a Jaccard score of 69.23% and a Dice score of 80.23%. Furthermore, using ConvMixer together with DenseNet121 in the analyses performed for CE-CD effectively increased the success. On the other hand, although the mutual exclusion problem was encountered, the proposed decoder was found to be effective.