Year: 2026
Type: preprint
Abstract: In this work, a hybrid segmentation architecture is proposed that integrates real, complex, and quantum computation within a U-Net framework. The model combines a ResNet-50 encoder for initial feature extraction, a complex-valued bottleneck convolutional block to preserve phase and amplitude information, and a quantum variational layer with eight qubits to encode global statistics derived from complex activations. The proposed model is evaluated on a polyp segmentation task using the PolipSet, Kvasir-SEG, and CVC-ClinicDB datasets, with direct baseline comparisons against a pretrained U-Net and SegFormer reported on PolipSet. Experimental results demonstrate that the hybrid complex--quantum U-Net achieves the best PolipSet performance among the compared models, with mean Intersection over Union (IoU) of 79.36\,\%, while also obtaining 78.29\,\% on Kvasir-SEG and 85.07\,\% on CVC-ClinicDB; the corresponding Dice coefficients are 88.38\,\%, 87.51\,\%, and 91.63\,\%, respectively. On PolipSet, these values outperform baselines such as Attention U-Net with ResNet-50 (IoU 75.93\,\%, Dice 82.94\,\%) and SegFormer (IoU 72.97\,\%, Dice 80.86\,\%). These findings highlight the potential of integrating complex-valued and quantum layers within deep learning pipelines for medical image analysis.
Source: Research Square
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214933222