Evaluation of Deep Convolutional Neural Networks for Detecting Nonpalpable Breast Abnormalities in Mammography
Imad Zyout · 2023
Advances in deep learning (DL) have renewed the motivation to build a computer-aided diagnosis (CAD) system for mammography as a feasible tool to diagnose breast cancer in its infancy. Current research reported the potential of the D L- based approach for characterizing malignancy of breast abnormalities. However, a preceding and desired feature of the complete mammography CAD, using DL, is the detection, as a first task, of breast abnormalities. This work focuses on developing multiclass DL-based algorithms to detect the most relevant and critical breast abnormalities (masses, clustered microcalcifications, architectural distortion), and the normal breast structure. For this purpose, several deep convolutional neural networks are trained and validated on a dataset of 1157 mammographic regions extracted from the Digital Database for Screening Mammography (DDSM), a public and common-use database. Preliminary results, in this work, have shown the impact of model selection but with insufficient performance in detecting mass and architectural distortion with most mispredicted representing incurred and so low precision and recall rates. To build a high-performance CADe, this work suggests adopting data-centric CAD in which data selection is an iterative and continuous process.