# Breast tumor segmentation and morphological feature-based classification in ultrasound using a two-stage U-net and SVM

## TL;DR

This paper introduces a two-step system using U-Net and SVM to detect and classify breast tumors in ultrasound images with high accuracy and interpretability.

## Contribution

A novel two-stage framework combining U-Net segmentation with SVM classification using morphological features for interpretable breast tumor diagnosis.

## Key findings

- The U-Net model achieved an average Mask IoU score of 91% for tumor segmentation.
- The SVM classifier reached 98.23% accuracy on the training set and 97.42% on the test set for benign vs. malignant classification.
- The framework preserves clinical interpretability by using handcrafted morphological features like circularity and solidity.

## Abstract

Breast cancer remains one of the most prevalent and life-threatening conditions among women worldwide, making early detection and accurate diagnosis essential. In this study, we present a two-stage computer-aided diagnosis framework designed for the automated analysis of breast ultrasound images.

The proposed system first employs a U-Net-based semantic segmentation model to detect and localize potential tumor regions. The model is trained and evaluated on a comprehensive dataset comprising normal, benign, and malignant cases. For each input image, the U-Net predicts a binary tumor mask; images with no detected tumor regions are classified as normal and excluded from further analysis. In the second stage, images identified as tumor-bearing undergo feature extraction to characterize the shape and morphology of the segmented tumor. Specifically, four handcrafted features—circularity, solidity, eccentricity, and extent—are computed from the predicted masks. These features are then used to train a support vector machine (SVM) classifier that distinguishes between benign and malignant tumors.

The segmentation model achieved an average Mask Intersection over Union% (Mask IoU) score of 91%, while the classification model reached an accuracy of 98.23% on the training set and 97.42% on the test set.

Unlike end-to-end deep learning approaches that often function as black boxes with limited clinical interpretability, our two-stage framework combines accurate deep learning-based segmentation with lightweight, handcrafted morphological feature classification using support vector machine. This design achieves high performance while preserving explainability through clinically meaningful shape descriptors, making it particularly suitable for real-world clinical deployment.