Chenhui88569/BCaDetection_WSI_CNN

WSIClassification_CNN

The project workflow

Three different models are generated

1. 4Conv_2FC + CWUR-HUP-TCGA cohort(Dataset1)
2. 4Conv_2FC + CWUR-HUP-TCGA-CINJ cohort(Dataset2)
3. 3Conv_2FC + CWUR-HUP-TCGA-CINJ cohort(Dataset2)

CNN model configuation

Two model architectures are used

1. 4Conv_2FC
   

2. 3Conv_2FC
   

Dataset constitution

Two different datasets are used

1. CWUR-HUP-TCGA cohort(Dataset1)
   CWUR-HUP-TCGA files used in training and validation set and CINJ used in test set.
   The numbers of image tiles extracted from four institutions are shown in the table

   	CWRU	HUP	TCGA	CINJ
   Training:positive	3000	2000	2000	0
   Training:negative	3000	2000	2000	0
   Training: total	6000	4000	4000	0
   Validation:positive	500	500	500	0
   Validation:negative	500	500	500	0
   Validation: total	1000	1000	1000	0
   Testing:positive	0	0	0	1500
   Testing:negative	0	0	0	1500
   Testing:total	0	0	0	3000

2. CWUR-HUP-TCGA-CINJ cohort(Dataset2)
   CWUR-HUP-TCGA-CINJ files used in training and validation set. CINJ used in test set

   	CWRU	HUP	TCGA	CINJ
   Training:positive	1750	1750	1750	1750
   Training:negative	1750	1750	1750	1750
   Total training	3500	3500	3500	3500
   Validation:positive	375	375	375	375
   Validation:negative	375	375	375	375
   Validation:total	750	750	750	750
   Testing:positive	0	0	0	1500
   Testing:negative	0	0	0	1500
   Testing: total	0	0	0	3000

Probability map

The Probability map is obtain by following the following procedure

1. Load model saved in disc
2. Regularly sample a input WSI and execute image preprocessing. The position of each individual tile is tracked
3. Predict the class of tach tile,obtain probablities
4. Reassemble the tiles. Only tissue tiles are given with probalities and the probablities of non-tissue tiles are zero.
5. Build heatmap