NYU open source breast cancer screening model trained on more than 200,000 mammographic exams


Breast cancer is the second leading cause of cancer death among women in the United States. It is estimated that in 2015, 232,000 women were diagnosed with the disease and about 40,000 died. And while diagnostic tests like mammography are widespread – in 2014, more than 39 million breast cancer screenings were done in the United States alone – they are not always reliable. About 10 to 15% of women who undergo a mammogram are asked to come back after an inconclusive analysis.

That's why researchers at New York University are studying an AI-based technique that promises much better accuracy than current tests. In an article recently published on Arxiv.org ("Deep Neural Networks Enhance the Performance of Radiologists in Breast Cancer Screening"), they describe a deep convolutional neural network – an algorithm class of 39, machine learning commonly used in the classification of images – which stuck an area. the ROC curve (AUC) of 0.895 to predict the presence of a cancerous breast tumor. In addition, they claim that, if the average of the probabilities of malignancy predicted by a radiologist from the results of the AI ​​system, the SSC is greater than that obtained by one or the other method .

The best-performing code and pre-workout models are available on Github.

"In this work, we propose a new neural network architecture … to effectively treat a large set of high resolution mammography mammography data with biopsy-proven labels," the authors of the paper explain. "We have experimentally demonstrated that our model is as accurate as an experienced radiologist and that it can improve the accuracy of radiologists' diagnoses if it is used as a second reader … The fact that the network found that previously invisible examinations with malignancies were similar, corroborates that our model has strong generalization capabilities. "

The team began by searching for a dataset of 229,426 digital mammography screenings (1,001,093 images) from 141,473 patients, each containing at least four images corresponding to the four views commonly used in mammographic examinations (right craniocaudal). left craniocaudal, oblique right medio-right). , and left mediolateral oblique). They extracted the labels from 5,832 examinations with at least one biopsy performed within 120 days of the screening mammogram, and then recruited a team of radiologists – all of whom received pathology reports to support – to indicate where biopsies were performed "at the pixel level". "

For each mammogram breast, researchers assigned two binary labels for a total of four labels: (1) absence or (2) presence of malignant signs and (3) absence or (4) presence of benign signs – that they used to form the convolution network. They also powered an auxiliary artificial intelligence system at the patch level with radiology-level pixel labels. They used it as well as the predictions of the main model to create thermal maps for exam images estimating the likelihood of malignant and benign results.

In experiments involving a series of 740 randomly selected biopsied and non-biopsied mammographic screening exams, despite the noise and ambiguity of some labels, the models yielded a prediction of 0.738 AUC and 0.895 AUC and 0.642 AUC. 0.779 AUC to predict malignant and benign tumors respectively, among patient populations.

To further validate the model, the authors of the paper conducted a study of 14 radiologists (12 assistants, a resident and a medical student). They were tasked with analyzing the test set consisting of 1,480 breasts. The AUCs ranged from 0.705 to 0.860, but when each was provided, the AI ​​system's predictions of cross-checking with theirs, the average AUC of the group increased to 0.891.

The researchers admit that the training dataset is relatively small and that, despite the architectural simplicity of the model, it would be impossible to form end-to-end most consumer materials. However, they say this is a promising step towards a generalizable model of cancer screening that could help clinicians make diagnostic decisions.

"[The results] We suggest that our network and radiologists have learned different aspects of the task and that our model can be an effective tool to provide a second reader for radiologists, "the researchers wrote. "Thanks to this contribution, research groups working to improve screening mammography, which may not have access to as extensive a training data set as ours, will be able to use directly our model in their research or use our pre-trained weights as initialization to form models with less data. By making our models public, we invite other groups to validate our results and test their robustness against changes in data distribution. "

NYU is not the only institution to apply breast cancer detection to AI. Last year, Google announced that a system developed through it, dubbed Lymph Node Assistant or LYNA, had reached an area under the receiver's operating characteristic of 99%. Baidu also claims to have designed an in-depth learning algorithm that surpasses human pathologists in its ability to identify breast cancer metastasis.


Source link