Researchers achieve accurate and non-invasive diagnosis of early lung cancer lymph node metastasis

Lung cancer is the cancer with the highest mortality rate in the world.
At present, the first choice of treatment for patients with early clinical lung cancer is lobectomy combined with systemic lymph node dissection, but for early stage lung cancer patients without lymph node metastasis, lymph node dissection will increase cancer recurrence and postoperative complications It may even lead to complications such as lymphedema, nerve damage, and pneumothorax
.

Therefore, accurate prediction of lymph node metastasis before surgery will effectively avoid unnecessary lymph node dissection, reduce the risk of recurrence and complications, and improve the quality of life of patients
.

At present, the preoperative prediction of early lung cancer lymph node metastasis status mainly depends on the doctor's judgment based on the experience of CT imaging.
This method is subjective, time-consuming, and has low accuracy (the average accuracy is about 0.
7)
.

The team developed a cross-modal information fusion neural network architecture to improve the diagnosis accuracy of early lung cancer lymph node metastasis in the early stage.
Although this method abandons the dependence of imaging omics on the manual delineation of lesions, it still requires clinicians to manually annotate lesion signs.
: Information such as burrs, lobes, pleural depression, primary focus density, etc.
, the model is not intelligent enough, and its clinical application is limited
.

In addition, the current stage of artificial intelligence technology in medical imaging intelligent diagnosis is poor in interpretability, which in turn reduces the confidence of clinicians in intelligent assisted diagnosis
.

In response to the above problems, the Gao Xin team, a researcher at the Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, based on previous research foundations, and aimed at the prediction of lung cancer lymph node metastasis, innovatively proposed a multi-scale, multi-task, multi-label classification neural network suitable for medical imaging.
Network model (Multi-scale, multi-task, and multi-label classification network, 3M-CN), the model architecture uses 3D DenseNet as the backbone of the model, extracts the three-dimensional CT features of lung nodules, and uses the multi-scale feature fusion module to deeply integrate the lungs The nodule images have different hierarchical features, and the multi-task segmentation module is used to guide the model to focus on the lesion area.
Finally, based on the multi-label classification task, the accurate prediction of the risk of lymph node metastasis and multiple signs and the search-based location and segmentation of the lesion area are realized simultaneously (Figure 1)
.

The study used the CT image data of 554 patients with early lung adenocarcinoma in two hospitals and the corresponding clinical data to train and verify the proposed network
.

The results show that the 3M-CN model proposed by the team can predict early lung adenocarcinoma lymph node metastasis with an accuracy of 0.
948
.

The advantage of the proposed method is that the model does not require any intervention by doctors, and only needs to use patient imaging data and clinical information to predict the risk of lymph node metastasis, and is fully automated and intelligent
.

At the same time, the model provides more semantic explanations related to lymph node metastasis, enhances the interpretability of deep learning models, and comprehensively improves clinicians' confidence in the results of the model, which is helpful for the clinical application of human-computer fusion
.

In addition, in order to further visualize the interpretability of the deep learning model, the team used visualization methods to quantify the correspondence between multi-scale features and the original image
.

Studies have shown that shallow neurons activate the marginal area of ​​the lung nodules, the representative sign is pleural depression; deep neurons activate more semantic sign areas, the representative signs are burrs, lobes and pleural depression (Figure 2 )
.

Related results were published in Computerized Medical Imaging and Graphics
.

The research was funded by the National Natural Science Foundation of China and other institutions
.

Figure 1 The proposed network 3D 3M-CN network architecture, MFF is a multi-scale feature fusion module, Segmentation Module is a segmentation module, FC is a fully connected layer, and RL is a detailed layer module to fuse clinical risk factors.
Figure 2 Three-dimensional multi-scale features The picture shows (a) the original CT image (the red arrow indicates the lobular sign; the green arrow indicates the pleural depression sign); (bi), (ci) and (di) are low, medium, and high-level feature activation maps, respectively; (b-ii) ), (c-ii) and (d-ii) are the low, medium, and high-level feature activation maps overlaid on the original CT images.
Source: Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences