Raú Ramos Pollán, Miguel Angel Guevara-López, Cesar Suárez Ortega, Guillermo Díaz-Herrero, Jose Miguel Franco-Valiente, Manuel Rubio del Solar, Naimy Gonzalez de Posada, Mario Augusto Pires Vaz, Joana Loureiro,Isabel Ramos
Medio de Publicación
Congreso: Ibergrid, Universidad Politécnica de Valencia
Tipo de publicación: Oral
Rationale and Objectives: This work aimed to evaluate a method to design mammography- based machine learning classiers (MLC) for breast cancer diagnosis. The main objectives were to validate the contents of the rst Portuguese \Breast Cancer Digital Repository" (BCDR) supporting the proposed method and obtain well performing classiers. Materials and Methods: We developed an image processing workstation and a data analysis toolkit suited for mining medical data, supported on distributed computing resources. From BCDR, 286 patient cases were extracted including digital content (mammography images) and associated metadata (clinical histories). Appropriated combinations of image processing techniques were applied to both craniocaudal (CC) and mediolateral oblique (MLO) images. Then, for each case (pathological lesion or normal tissue) a region of interest (ROI) was identied and manually segmented by specialized radiologists in one or both images. Several datasets were built combining features vectors computed for segmented ROIs and, nally, a massive exploration of MLC was made to obtain automatic BI-RADs diagnosis for each dataset. Results: BCDR includes critical information to make breast cancer studies. Suitable combinations of image processing techniques were identied to reduce image artifacts and improve mammography details. ANN and SVM classiers achieving ROC Az between 0.661 and 0.996 were
obtained for the dierent datasets constructed. The proposed method covers the complete MLC development cycle for BI-RADS classication.
Conclusion: Mammography-based MLC built from datasets including features vectors representing the same ROIs identied in both CC and MLO images revealed to be advantageous for dierentiating malignant and benign lesions.