Work

Hybrid MobileNetV2–ELM Framework for Multiclass Breast Cancer Classification in Digital Breast Tomosynthesis Images

Year: 2026

Type: conference-paper

Abstract: Digital Breast Tomosynthesis (DBT) provides an extensive view of breast tissue for breast cancer screening. Nevertheless, the volumetric nature of DBT requires high computational power and memory capacity. To overcome this limitation, this work proposes a lightweight hybrid MobileNetV2-Extreme Learning Machine (ELM) framework for four-class DBT image classification using two-dimensional image representation. The dataset comprises 2,000 balanced DBT images in four categories, namely actionable, benign, malignant, and normal. The images were resized to 200 x 200 x 3 and split into training, validation, and testing data with a ratio of 70:15:15. We studied three classification scenarios: base MobileNetV2 with softmax, fine-tuned MobileNetV2 with softmax, and fine-tuned MobileNetV2 feature extraction with ELM. The proposed MobileNetV2-ELM model obtained the highest accuracy of 76.00% with a precision of 75.82%, recall of 76.00%, and F1-score of 75.54%. The results show that ELM can slightly improve the classification performance when used on deep features extracted from MobileNetV2. This study shows the potential of lightweight hybrid deep feature learning for multiclass DBT image classification under limited computational resources.

Source: International Conference on Artificial Intelligence, Robotics, Signal and Image Processing (AIRoSIP)

DOI: 10.1109/airosip71802.2026.11741523