Author Affiliations
1Key Laboratory of Pattern Recognition and Intelligent Image Processing, College of Information Engineering, Inner Mongolia University of Science and Technology, Baotou, Inner Mongolia 0 14010, China2Institute of Information Engineering, Inner Mongolia University of Technology, Hohhot, Inner Mongolia 0 10051, China3College of Computer Engineering and Science, Shanghai University, Shanghai 200444, China4College of Information Science and Technology, Dalian Maritime University, Dalian, Liaoning 116026, Chinashow less
Fig. 1. Flow chart of our algorithm
Fig. 2. Working principle of OctConv transition layer
Fig. 3. Convolution with different structures. (a) Traditional convolution; (b) HetConv
Fig. 4. Flow chart of HetConv algorithm
Fig. 5. Structure of ResNeXt module
Fig. 6. Image of the Benign class in the verification set. (a) Whole image; (b) small patches
Fig. 7. Principle of the majority voting algorithm
Fig. 8. Training accuracy and verification accuracy of the ResNeXt model
Fig. 9. Image of partially judged wrong. (a) Invasive; (b) InSitu1; (c) InSitu2
Fig. 10. Training accuracy and verification accuracy of the ResNeXt+OctConv model
Fig. 11. Image of the Normal class
Fig. 12. Training accuracy and verification accuracy of the ResNeXt+OctConv+HetConv model
| Benign | InSitu | Invasive | Normal |
---|
Benign | 18 | 1 | 1 | 3 | InSitu | 1 | 18 | 4 | 0 | Invasive | 0 | 1 | 14 | 1 | Normal | 1 | 0 | 1 | 16 |
|
Table 1. Image-level confusion matrix of ResNeXt
| Benign | InSitu | Invasive | Normal |
---|
Benign | 18 | 1 | 0 | 1 | InSitu | 1 | 18 | 3 | 0 | Invasive | 0 | 0 | 17 | 0 | Normal | 1 | 1 | 0 | 19 |
|
Table 2. Image-level confusion matrix of ResNeXt+OctConv model
| Benign | InSitu | Invasive | Normal |
---|
Benign | 18 | 1 | 0 | 1 | InSitu | 0 | 18 | 2 | 0 | Invasive | 0 | 0 | 18 | 0 | Normal | 2 | 1 | 0 | 19 |
|
Table 3. Image-level confusion matrix of ResNeXt +OctConv+HetConv model
| Benign | InSitu | Invasive | Normal |
---|
Benign | 23 | 1 | 1 | 4 | InSitu | 1 | 20 | 2 | 1 | Invasive | 0 | 1 | 22 | 0 | Normal | 1 | 3 | 0 | 20 |
|
Table 4. Image level confusion matrix of Ref. [4]
Method | Recall | Precision | Accuracy |
---|
Ourmethod | Benign | 90.00 | 90.00 | 91.25 | InSitu | 90.00 | 90.00 | Invasive | 90.00 | 100.00 | Normal | 95.00 | 86.36 | Ref. [4] | Benign | 92.00 | 79.31 | 85.00 | InSitu | 80.00 | 83.33 | Invasive | 88.00 | 95.65 | Normal | 80.00 | 83.33 |
|
Table 5. Recall,precision and accuracy of two methods unit: %
Method | ResNeXt | ResNeXt+OctConv | ResNeXt+OctConv+HetConv P=2(P=4) | Ref.[4] |
---|
Patch-accuracy | 71.92 | 81.73 | 83.04(78.12) | 79.00 | Image-accuracy | 82.50 | 90.00 | 91.25(88.75) | 85.00 |
|
Table 6. Recognition rate of different models unit: %
Method | Accuracy |
---|
Traditional machine learning[1] | 80.00-85.00 | AlexNet[2] | 89.60 | CNN+SVM[3] | 77.80 | Inception-Transfer learning[4] | 85.00 | LightGBM[5] | 87.20 | Hierarchical ResNeXt[6] | 99.00 | The contestants (ICIAR2018) | 80.00-91.00 | Our method | 91.25 |
|
Table 7. Experimental results obtained by different methods unit: %