A Hybrid Deep Learning Model for Tumors and Organs segmentation in Cervical Cancer using Kolmogorov-Arnold Convolutions

Cervical cancer is the second highest incident cancer among women. Radiation therapy is a common non-surgical treatment. Proper segmentation of cervical tumor and healthy tissue, such as the bladder and rectum, is critical for complete cancer elimination, minimal healthy tissue toxicity, and a positive patient outcome. Deep learning via convolutional neural network (CNN) models has become a breakthrough in automating and standardizing tumor segmentation for more accurate treatment planning. Models come in a diverse array of architecture, each bringing their own strengths and weaknesses. Hybrid models integrate the predictive power of multiple models, providing more accurate segmentation. In this study, we developed two hybrid-based models for cervical cancer using Kolmogorov-Arnold Convolutions (Hybrid-ConvKAN) and multilayer perceptrons (Hybrid-MLP) that combines the outputs of Deeplabv3+, 3D Unet and a custom 2D model. The two models automatically segment 11 organs (cervix-uterus, vagina, bladder, rectum, sigmoid, bowel bag, spinal cord, femoral heads, kidneys), CTVp (uterus and parametria) and CTVn (nodal) from CT images. The 11 organs and CTVs were delineated from 157 patients with 223 CT scans divided into 149 training, 36 validation and 38 test CTs. An additional 30 external CT scans were used for testing our proposed model. The two models use Adam optimizer and a custom loss function (cross entropy and dice loss). Performance was measured using the dice similarity coefficient (DSC), surface DSC at 2 mm (sDSC2mm), precision, and recall. From the experiments, Hybrid-ConvKAN segmented the 11 OARs and CTVs more accurately compared with the individual models and Hybrid-MLP. The average DSC for Hybrid-ConvKAN was 0.91 (range: 0.82-0.94) for organs, 0.83 for CTVp (range: 0.79-0.87) and 0.81 for CTVn (range: 0.77-0.84). The comparison methods were ranged from 0.73 and 0.92 for organs, and 0.64 and 0.84 for CTVs. The proposed Hybrid-ConvKAN model is promising for implementing dose stratification during the treatment.