Arsenii Litvinov’s research while affiliated with Institute for System Programming, Russian Academy of Sciences and other places

In this study, we introduce a novel method, called MamT$^4$, which is used for simultaneous analysis of four mammography images. A decision is made based on one image of a breast, with attention also devoted to three additional images: another view of the same breast and two images of the other breast. This approach enables the algorithm to closely replicate the practice of a radiologist who reviews the entire set of mammograms for a patient. Furthermore, this paper emphasizes the preprocessing of images, specifically proposing a cropping model (U-Net based on ResNet-34) to help the method remove image artifacts and focus on the breast region. To the best of our knowledge, this study is the first to achieve a ROC-AUC of 84.0 $\pm$ 1.7 and an F1 score of 56.0 $\pm$ 1.3 on an independent test dataset of Vietnam digital mammography (VinDr-Mammo), which is preprocessed with the cropping model.

BACKGROUND: Breast cancer is one of the leading causes of cancer-related mortality in women [1]. Regular mass screening with mammography plays a critical role in the early detection of changes in breast tissue. However, the early stages of pathology often go undetected and are difficult to diagnose [2]. Despite the effectiveness of mammography in reducing breast cancer mortality, manual image analysis can be time consuming and labor intensive. Therefore, attempts to automate this process, for example using computer-aided diagnosis systems, are relevant [3]. In recent years, however, solutions based on neural networks have gained increasing interest, especially in biology and medicine [4-6]. Technological advances using artificial intelligence have already demonstrated their effectiveness in pathology detection [7, 8]. AIM: The study aimed to develop an automated solution to detect breast cancer on mammograms. MATERIALS AND METHODS: The solution is implemented as follows: a deep neural network-based tool has been developed to obtain the probability of malignancy from the input image. A combined dataset from public datasets such as MIAS, CBIS-DDSM, INbreast, CMMD, KAU-BCMD, and VinDr-Mammo [9–14] was used to train the model. RESULTS: The classification model, based on the EfficientNet-B3 architecture, achieved an area under the ROC curve of 0.95, a sensitivity of 0.88, and a specificity of 0.9 when tested on a sample from the combined dataset. The model’s high generalization ability, which is another advantage, was demonstrated by its ability to perform well on images from different datasets with varying data quality and acquisition regions. Furthermore, techniques such as image pre-cropping and augmentations during training were used to enhance the model's performance. CONCLUSIONS: The experimental results demonstrated that the model is capable of accurately detecting malignancies with a high degree of confidence. The obtained high-quality metrics offer a significant potential for implementing this method in automated diagnostics, for instance, as an additional opinion for medical specialists.