[BELLE HISTOIRE] Using AI to help detect breast cancer

According to Santé publique France, breast cancer is the leading cause of cancer-related death in women. This is why the French government, as part of a national screening program, recommends that all women aged between 50 and 74 undergo screening every two years. The examination generally consists of a 2D mammogram: the breast is placed between an X-ray source and a detector to obtain a grayscale image, reflecting the presence of various materials - fat, mammary gland or tumors.

Breast tomosynthesis: greater precision but limited image quality

But since the 2010s, a new screening method has emerged: 3D mammography, or breast tomosynthesis, an imaging modality akin to CT scanning.

In a CT scanner, the X-ray generator and detector rotate 360° around a patient lying down. A series of images is acquired during this rotation, which are then used to reconstruct a 3D volume of the area examined. Based on a similar principle, breast tomosynthesis provides more precise images, thus facilitating diagnosis by practitioners. The use of this technique is showing promising results, particularly when coupled with conventional 2D mammography.

In the case of tomosynthesis, however, the method comes up against inherent limitations. " Unlike a CT scan, it is not possible to rotate 360° around the patient," notes Arnaud Quillent, image processing research engineer at GE HealthCare. " A complete rotation would mean moving the patient, because, unlike her position in a scanner, she is standing upright. This movement would make 3D reconstruction much more complex. And generally speaking, the recurrence of screening examinations means that we need to limit the dose of X-rays to which an individual is subjected. "

As a result, current tomosynthesis equipment only covers an angle of around 15° to 40°, a far cry from the 360° of the scanner. This limitation significantly impairs the quality of the 3D reconstruction produced from the images collected: due to the lack of information, the breast elements visible in the reconstructed volume "smear" and blend in with the surrounding structures. These degradations, known as "artifacts", can reduce the visibility of tumors within the 3D volume.

Fake breasts to train AI model to erase artifacts

This is why Arnaud, as part of his thesis involving the Laboratoire Traitement et Communication de l'Information (LTCI) at Télécom Paris - part of the Carnot TSN institute - and GE HealthCare, set out to develop a method for reducing these artifacts. His idea was to use an artificial intelligence model, more specifically deep learning, that would learn to detect and erase the artifacts present in breast tomosynthesis images.

The first step was to train the neural network on which the deep learning algorithm is based. It had to be presented with a series of images to compare with a reference dataset, i.e. a tomosynthesis image showing no artifacts. Problem: such a resource doesn't exist, due to the limited rotation of all existing devices - which is, incidentally, the raison d'être of the thesis.

So, how to feed the AI algorithm in its learning phase? " We had to generate synthetic breast models ourselves, based on images from other screening techniques," reveals Arnaud. " Breast scanning, which is less widespread than tomosynthesis due to a number of drawbacks, can generate images free of the artifacts we're looking to correct. We then used these images to create 3D models of false breasts, identifying the tissues: fat, mammary gland, skin...". This work required a crucial adaptation phase, as the quality of the scanner differs from that of tomosynthesis.

Virtual mammograms for artificial breasts

The researcher was able to produce several hundred models of artificial breasts. A convincing result, but insufficient to train a neural network. To make up for this shortfall, Arnaud proceeded by "data augmentation": he varied certain parameters of the training images to obtain new ones and thus increase their number.

Next, the artificial breasts underwent a mammography examination... virtually. The researcher used a computer to simulate acquisitions by a tomosynthesis device with a limited angle of rotation. Thanks to this operation, the neural network had all the elements it needed to learn: images with artifacts to compare with "perfect" breast models.

AI capable of self-assessing its reliability

In this way, the deep learning model learned to recognize artifacts and make them disappear. As a result, it was able to improve the quality of breast tomosynthesis images. But how much can we trust the reconstructions it produces? " This is a key question, especially in the medical field," confirms Arnaud. " And to answer it, we used... the same AI model. In fact, it doesn't just reconstruct enhanced images: it also produces an "uncertainty map". Within this, for each pixel of the image, a level of doubt is calculated, with a low value translating into a high degree of certainty, and vice versa. " A way of informing the user about the reliability of the information generated.

But how can AI itself assess the relevance of its own productions? " During the training phase, the neural network also learns that certain areas of the breast may be more or less difficult to reconstruct," explains Arnaud. " And we can guide it in its learning by comparing the reliability it has estimated with its actual error level, i.e. the measured difference between the reconstructed image and the reference one. "

Ultimately, for each input, the deep learning model delivers two outputs: a reconstruction of the breast image and an uncertainty map. This dual mission required the neural network to be optimized, given that the two tasks carried out in parallel can overlap. This balancing act was in addition to the challenge of exploring a field - AI reconstruction of mammography images - that has yet to be tackled in the scientific literature, due to the small amount of data available.

The deep learning model put to the test in clinical cases

While the model we have developed has produced satisfactory results on a test corpus of computer-generated images, there are still a number of stages to go through before it can be used by practitioners. Firstly, it needs to confirm its relevance on higher-quality images. Indeed, to date, it has only been confronted with image resolutions lower than those of tomosynthesis devices, due to the memory constraints imposed by the hardware available at the time of the thesis.

Above all, a major next step would be to put the deep learning model to the test on real images from clinical trials. The aim would then be to demonstrate that the results obtained on synthetic data are also valid for real patient cases. This would represent a significant step forward towards better breast cancer detection.