First, the model
The core model uses now Andrew Ng's ChexNet model that is DenseNet 121 layers. Although the diagnosis model in the given data set is better than the results of four diagnostic imaging department, and Andrew Ng also said Should radiologists be worried about their jobs Breaking news:? We can now diagnose pneumonia from chest X-Rays better than radiologists. but Diss attracted a large, after all, this conclusion would be under a little early in the image circle. However, an individual or a team standing in artificial intelligence.
Our production environment before this model, joined the orthotopic model and image classification of other sites. Training environment, it is also added to the Unet lung segmentation model, before image processing, for example, no image (window center \ window width) of the window width is automatically calculated.
Second, the key code interpretation model
1 random amplification shear treatment
normalize = transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
transformList = []
transformList.append(transforms.RandomResizedCrop(transCrop))#随机放大剪切
transformList.append(transforms.RandomHorizontalFlip())
transformList.append(transforms.ToTensor())
transformList.append(normalize)
ChexNet method employed, the DR is the original image to reduce the image size of 3000X3000 1024. Only reduced three times the size than the original, texture detail loss is relatively small, and computer perception of textures certainly stronger than the human eye can distinguish more subtle texture. In addition, most now display the DR diagnosis, and only 1M, a lot of time, to observe teachers do not magnify the image. So, these two reasons, the reduced 3 times the original size, is acceptable. When we deal with this part of the image, the first is to image cropping to size after about unilateral 2800, this scaled to 1024, it retains more information map.
2 ChexNet image processing method RandomResizedCrop 1024, that is on the original, random scaling, and wherein the random intercept image size of 224. Here developers may be in doubt, at some point, did not intercept the location of the lesion, such as a picture or as a positive sample, the model training. Barbara reason for this is that in the negative sample, also intercepted a similar image area, has been identified as negative samples, after so many times of training, this characteristic of the image must have been suppressed, Barbara lesions special clinics relatively increased.
3 and ChexNet different is that we have to lung region image is divided, before training with the segmented image. Segmentation is used to model the short added in the process ResNet UNet basis. Under normal circumstances, the lung segmentation results is very good, as shown below. But for some lung lesions, the effect can not be satisfied in any case, as shown in FIG. 2, right upper lung produced becomes solid, but not recognized segmentation algorithm. Therefore, we in the segmented image and made some conventional algorithms, and finally the resulting image, shown in Figure 3.
Figure 1
Figure 2
Figure 3
Third, the training results
We from the database, marked about 11,000 tuberculosis, 11,000 negative samples. FIG ROI below
Fourth, the project achieve results
After selecting the threshold is 0.42, and then sent to the front end positioned FIGS H5, js use of the thermal and FIG dicom grayscale merge, shown below
Corresponding teacher follows
{"finding":"双侧胸廓大致对称。右上肺可见斑点及索条状密度增高影，边界清晰。余肺内未见其它异常阴影。肺门形态、大小及位置未见异常。纵隔无增宽表现，位置居中无移位，心影不大，形态未见异常。双侧膈肌光整，肋膈角锐利。片内所示胸廓骨骼及胸壁软组织无异常表现。","result":"右上肺结核，病灶基本纤维化、钙化。"}
