Mammography is known as one of the best forms to screen possible breast cancer in women, and recently deep learning models have been developed to assist the radiologist in the diagnosis. However, their lack of interpr...
Mammography is known as one of the best forms to screen possible breast cancer in women, and recently deep learning models have been developed to assist the radiologist in the diagnosis. However, their lack of interpretability has become a significant drawback to their extended use in clinical practice. This paper introduces a novel approach for detecting and localising pathological findings in mammography exams through the use of a EfficientNet-based deep learning model. The model is trained using cropped segments of labelled pathological findings from Vindr Mammography Dataset. Achieving an average F1-score of 72.7 %, and reaching on mass and suspicious calcifications an F1-Score of 79.9 % and 84.5 % respectively. Using this classifier we propose a method to visualise from local information the regions of interest where pathological findings could be present on the complete image. Plus, we describe the limitations regarding area coverage of these patches on the model’s capability of generalization and certainty on its predictions, explaining its functionality.
暂无评论