Breast cancer is one of the most common cancers in women, with more than 1,300,000 cases and 450,000 deaths each year worldwide. In this context, recent studies showed that early breast cancer detection, along with suitable treatment, could significantly reduce breast cancer death rates in the long term. X-ray mammography is still the instrument of choice in breast cancer screening. In this context, the false-positive and false-negative rates commonly achieved by radiologists are extremely arduous to estimate and control although some authors have estimated figures of up to 20% of total diagnoses or more. The introduction of novel artificial intelligence (AI) technologies applied to the diagnosis and, possibly, prognosis of breast cancer could revolutionize the current status of the management of the breast cancer patient by assisting the radiologist in clinical image interpretation. Lately, a breakthrough in the AI field has been brought about by the introduction of deep learning techniques in general and of convolutional neural networks in particular. Such techniques require no a priori feature space definition from the operator and are able to achieve classification performances which can even surpass human experts. In this paper, we design and validate an ad hoc CNN architecture specialized in breast lesion classification from imaging data only. We explore a total of 260 model architectures in a train-validation-test split in order to propose a model selection criterion which can pose the emphasis on reducing false negatives while still retaining acceptable accuracy. We achieve an area under the receiver operatic characteristics curve of 0.785 (accuracy 71.19%) on the test set, demonstrating how an ad hoc random initialization architecture can and should be fine tuned to a specific problem, especially in biomedical applications.
Current Status and Problems of Mammographic Screening for Breast Cancer in Japan
