Transformer-Convolution Pyramid Network for Multi-Class Kidney, Tumor, and Cyst Segmentation

Abstract

Background/Objectives: Chronic kidney diseases are fatal disorders, and early treatment is crucial for reducing mortality. Segmenting kidneys, tumors, and cysts from computed tomography scans has a significant role in assessing these diseases’ progression, determining their severity, and planning surgeries. However, accurate segmentation is challenging for many reasons. The objective of this study is to develop a hybrid network that can overcome these challenges. Methods: The proposed network was designed using (i) a patch-joining and swin transformer block to capture global features efficiently, (ii) a Dynamic Channel Attention block to emphasize significant feature locations with the help of channel-wise attention, (iii) a block to perform complementary fusion of features coming from different layers, (iv) an edge-detection block to improve the performance by paying extra attention to edge features, and (v) an efficient activation function to overcome dying neuron and bias-shifting problems caused by the commonly used rectified linear unit. Extensive experiments were conducted using images from KiTS-2023 and Kaggle databases. Also, state-of-the-art techniques were applied to identical datasets and evaluated using identical metrics for fair comparisons. In addition to the commonly used Dice Similarity Coefficient (DSC), the Normalized Surface Distance (NSD) was used for evaluation. Results: The proposed method yields DSC/NSD scores of 98.35 ± 0.65%/97.62 ± 0.55%, 94.88 ± 0.57%/93.84 ± 0.50%, and 81.99 ± 0.45%/80.93 ± 0.40% for kidneys, tumors, and cysts, respectively. Conclusions: The results indicate the superiority of the proposed approach in segmenting kidneys, tumors, and cysts. The proposed method has the potential to support medical experts by providing quantitative and objective segmentation results.