A novel classification method of lymph node metastasis in colorectal cancer

Figures & data

Figure 1. CRC LN. The top row are negative, the bottom row is positive

Figure 2. Architecture of our method

Table 1. Pseudo-code of the method algorithm

Algorithm: Method algorithm

1 : Input(h, w, c); 2 : F output, F total, f 1, f 2 … … f n ; 3 : f 1 (Input(h, w, c)) ; 4 : f 2 ; (f1) ; 5 : f 3 (f1 + f2); … … 6 : f n (f 1 + f 2 + … … +f n-1) ;

7 : F total (f 1 + f 2 + … … + f n) ; 8 : Δ ‘+’ indicates concatenate operation 9: Δ ‘’ indicates convolution operation

10: F output [∑ (F total) ^p] ^1/p ;

11 : return F output

Figure 3. Structure of AlexNet-A

Figure 4. Structure of AlexNet-B

Figure 5. Structure of AlexNet-C

Figure 6. Structure of AlexNet-D

Figure 7. Structure of AlexNet-E

Figure 8. Relationship between accuracy and dimension

Table 2. Accuracy of five structure

Name	AlexNet-A	AlexNet-B	AlexNet-C	AlexNet-D	AlexNet-E
Accuracy	0.7598	0.7725	0.7968	0.8088	0.8156

Figure 9. Accuracy curve of six methods on CRC LNM classification

Figure 10. ROC curve of six methods on CRC LNM classification

Table 3. Classification result on crc lnm metastasis classification

Method	Sensitivity	Specificity	PPV	NPV	Accuracy	AUC
AlexNet model	0.6708	0.6711	0.6714	0.6706	0.6716	0.7696
AlexNet-pretrained model	0.8004	0.7997	0.7992	0.8009	0.7583	0.7941
CNN-AlexNet with SVM	0.7015	0.7015	0.7015	0.7015	0.6920	0.7702
DDC	0.5	0.3478	0.7196	0.1720	0.4670	–
DAN	0.4393	0.4815	0.8444	0.1182	0.4834	–
ResNet152	0.6801	0.6801	0.6801	0.6801	0.6801	0.7327
DenseNet161	0.625	0.625	0.625	0.625	0.625	0.6281
Ours	0.8732	0.8741	0.8746	0.8728	0.8358	0.8569

‘ – ’ indicates the value of AUC less than 0.5

Figure 11. CRC LN classification heat-map. Left is the original image; the middle is the feature heat-map; right is the superimposed image

Table 4. Classification result of four radiologists and our method

Method	Sensitivity	Specificity	PPV	NPV	Accuracy	AUC
Radiologist1	0.5443	0.7091	0.6467	0.6142	0.6279	0.6263
Radiologist2	0.6823	0.6047	0.6278	0.6616	0.6432	0.6433
Radiologist3	0.6986	0.6445	0.6573	0.6613	0.6717	0.6711
Radiologist4	0.7239	0.6174	0.6481	0.6896	0.6694	0.6699
Ours	0.8732	0.8741	0.8746	0.8728	0.8358	0.85697

Figure 12. ROC curve of radiologists and our method on CRC LNM classification