Segmentação e classificação experimental de lesões mamográficas com redes neurais convolucionais

Abstract

This work aimed to develop a Python application for the segmentation and experimental classification of mammographic lesions using CNNs. The CBIS-DDSM database was used, containing mammographic images, pathological information, and ROI annotations. The methodology was divided into two main stages: segmentation of suspicious regions using U-Net and classification of lesions as benign or malignant using Transfer Learning-based models. In the segmentation stage, mammograms associated with their respective ROI masks were used, allowing supervised training of the U-Net model. After mask prediction, a post-processing step was applied to remove noise, fill gaps, and visually highlight the suspicious region. In the classification stage, different approaches were tested, including an initial CNN, MobileNetV2, and EfficientNetB0. Adjusting the decision threshold for the malignant class increased the recall of this class, prioritizing the identification of malignant cases. The results showed that segmentation still presents limitations, with a validation Dice of 0.2459 and a validation IoU of 0.1453. In classification, the EfficientNetB0 model with a threshold of 0.20 achieved a recall of 0.71 for the malignant class, although with an increase in false positives. The developed application integrated segmentation, post-processing, classification, and automatic saving of results. It is concluded that the system has an academic and experimental purpose and is not suitable for clinical use, but it demonstrates the potential of CNNs in mammographic image processing.