PHnet: A Hybrid CNN-MLP Framework for Efficient and Accurate Uterine Tumor Segmentation in 3D Medical Imaging

Abstract

Uterine tumors pose a major health risk, and for effective treatment planning, it is essential to identify them quickly, classify them correctly, and segment them accurately. Traditional diagnostic techniques that depend on the manual interpretation of medical images are frequently slow and prone to inter-observer discrepancy [1]. The precision and efficiency of automated uterine tumor analysis have been significantly enhanced by recent developments in deep learning. This research offers a complete strategy that combines segmentation and classification through sophisticated neural architectures. A new hybrid model called PHNet is proposed for segmentation. It combines 2D and 3D Convolutional Neural Networks (CNNs) with a Multi-Layer Permute Perceptron (MLPP) to effectively capture local features and global context, tackling issues with anisotropic volumetric data. An improved Vision Transformer (ViT)-based model is created for classification, featuring a novel relative positional encoding method and residual MLP blocks to enhance spatial awareness and convergence rate. Image preprocessing methods like Homomorphic Filtering, CLAHE, and Unsharp Masking are used to mitigate the small dataset and improve model generalization. Experimental results on augmented uterine tumor datasets show that the method outperforms both segmentation and classification tasks, with significant gains in precision, recall, and accuracy. By integrating explainable AI and graph-based techniques like Deep LOGISMOS, the combined framework also promotes clinical interpretability, providing a viable, expandable answer for practical diagnostic workflows.