Abstract:

Early diagnosis of breast cancer remains one of the most critical factors for a successful 
treatment outcome, with mammography continuing to be the "gold standard" for screening. 
However, the visual analysis of mammographic images is a complex process prone to subjective 
errors and radiologist fatigue. This study explores the integration of Deep Learning methods to 
automate the diagnostic process and enhance accuracy in pathology detection. The primary 
objective is to create a clinical decision support model that acts as a "second opinion," classifying 
early-stage anomalies using the annotated CBIS-DDSM dataset. 
The research methodology involves constructing a specialized data preprocessing pipeline that 
transforms raw DICOM images into 16-bit PNG format, preserving the informational depth 
necessary for precise analysis. A binary classification strategy (masses vs. calcifications) was 
applied using a Convolutional Neural Network (CNN) with a ResNet-101 architecture and transfer 
learning. A custom data splitting scheme (80% training, 10% validation, 10% testing) was 
introduced, allowing for stricter control over the training process and the prevention of overfitting. 
The results demonstrate that combining proper preprocessing (cropping methods to isolate 
significant visual information and normalization) with powerful computer vision architectures leads 
to the creation of a robust model suitable for recognizing image variations. The developed system 
does not aim to replace the physician but to optimize the clinical workflow by filtering and 
categorizing potential lesions with high accuracy. This research highlights the potential of 
bioinformatics to transform massive medical data into clear diagnostic indicators, reducing 
analysis time and the risk of missing microscopic anomalies. 

Keywords: intelligent systems, neural networks, recognition, diagnosis, mammography 

Download the full article here: Download 2016