U-shaped Network Utilizing Tokenized Mlp and Strong Convolution for Ultrasonic Medical Image Segmentation

When diagnosing ultrasonic medical images, if the lesion area can be automatically and accurately segmented, it will assist physicians in objectively analyzing and making prompt and precise diagnoses. The U-Net architecture and its subsequent developments have demonstrated significant achievements in the domain of medical picture segmentation. Nevertheless, the U-Net encoders face challenges in efficiently extracting global contextual information as a result of the intrinsic local characteristics of conventional convolution operations. Furthermore, a simplistic skip connection fails to capture significant features. This paper introduces a new U-shaped network called MMUnet, which aims to overcome the limitations listed above. To do this, MMUnet combines a robust convolution-based hybrid medical imaging segmentation network with a tokenized MLP. The MMUnet utilises a combination of hybrid convolutions and coordinate attention gates. The MLPConvMixer module facilitates the extraction of global contextual information by merging features from distant spatial locations. Additionally, the coordinate attention gates emphasize valuable features and enable efficient connection jumps. We evaluated the effectiveness of the proposed methods using thyroid ultrasound image datasets and breast ultrasound datasets. The MMUnet model produced mean Crossover over Union (IoU) values of 67.59% and 69.29%, along with 79.65% and 81.85% F1 scores.