Robust multi-modal ovarian tumor segmentation using PVTv2 and attention-guided deformable convolutions

Accurate ovarian tumor segmentation on ultrasound is important for quantitative assessment and can reduce subjective variability in clinical interpretation, yet remains difficult due to speckle noise, low contrast, and irregular lesion boundaries. We propose PVT-CBAM-CONV (Pyramid Vision Transformer with Convolutional Block Attention Module and Deformable Convolution), a hybrid segmentation framework that combines a pyramid Transformer encoder for multi-scale global context, lightweight attention-based feature refinement, and deformable convolutions to better adapt to complex tumor contours. To improve robustness under severe foreground–background imbalance, we employ a hybrid loss combining weighted BCE (Binary Cross-Entropy) and Dice with training-stabilization safeguards. Experiments on the public MMOTU (Multi-Modality Ovarian Tumor Ultrasound) dataset (OTU_2D and OTU_CEUS) show that PVT-CBAM-CONV consistently outperforms representative CNN (Convolutional Neural Network)- and Transformer-based baselines, achieving 75.74% IoU (Intersection over Union) on OTU_2D and 81.12% IoU on OTU_CEUS. With 27.8M parameters and 15–18 FPS (Frames Per Second) inference at 512×512 on an RTX 3070, the proposed method demonstrates a practical balance between accuracy and efficiency, supporting its potential use as an assistive tool for ovarian tumor assessment across routine ultrasound and contrast-enhanced ultrasound.