Radiomics-integrated machine learning framework for quantitative breast cancer diagnosis

TL;DR

This paper presents an automated machine learning framework combining deep learning and radiomics to improve breast cancer diagnosis from ultrasound scans.

Contribution

The novel contribution is an integrated UNet-Radiomics-ML framework for automated tumor segmentation and classification in breast ultrasound imaging.

Findings

The framework achieves a mean IoU of 0.94231 for tumor segmentation.

It attains 97.8% classification accuracy for benign and malignant tumors.

The system reduces human intervention and accelerates the diagnostic process.

Abstract

Breast cancer is one of the most common cancer types affecting women worldwide, and its early detection is crucial for effective treatment. The proposed study offers an automated pipeline that uses deep learning, radiomics, and machine learning to segment and classify breast tumors. The pipeline takes ultrasound scans as input, segments the tumor using UNet, and uses the predicted segment to compute the radiomic features, which are then given as input to the machine learning models for classification.

In the first phase of the proposed study, the ultrasound scans and the ground truth masks in the BUSI dataset, obtained by the radiologist, are used to extract the radiomic features, followed by training the machine learning (ML) models. These results are used as benchmarks to evaluate the performance and efficacy of the proposed segmentation model. The use of radiomics bridges the gap between medical imaging and quantitative analysis. The features extracted by the proposed system are shape-based and therefore provide morphological information about tumor cells. In the second phase, an automated deep learning pipeline is proposed, in which the UNet is trained to effectively segment tumorous regions in ultrasound scans. Once the segmentation mask is obtained, it is used to compute radiomic features, and the ML algorithms are trained to classify the tumor as benign or malignant.

The proposed UNet-Radiomics-ML framework achieves performance when compared to the benchmark results obtained in the first phase. The proposed framework achieves a mean IoU of 0.94231 and a classification accuracy of 97.8% for the testing data.

The results obtained from this study truly have the potential to positively impact the diagnosis of breast cancer. The use of automated models like UNet for segmentation, radiomics for feature extraction, and ML algorithms for classification will help reduce human intervention and narrow the diagnostic process, yielding quick results.

Figures