gatsby2016 / DLforWSI

An effective deep learning classification framework for whole slide images.

projects / aim

ccRCC tumor detection by DL sliding widows on whole slides images.

data: slides and patches

In total of 5 slides 4 for training and 1 one for validation

patches distribution	non	tumor
training	23824	40946
validation	10136	12605

The patch size is 512*512 at 20x magnificaiton.

with samples selection manually slightly. Mainly for whole white patches.

No stain normalizaiton, no samples agumentation

2019/11/2 update

methods

• DL resnet34 + weighted cross entory loss for 45 epochs with lr 0.001 on pretrained ImageNet.
• weights are: [0.7, 0.3]

results on epoch 45

ACC: 0.8814, take about 302 seconds

Confusion Matrix	non	tumor	total
real non	7451	2685	10136
real tumor	11	12594	12605

• acc = 7451+ 12594 / (10136+12605)= 0.88144761
• sen = recall = 12594/ 12605 = 0.999127
• spc = 7451 / 10136 = 0.7351
• precision = 12594 / (12594+2685) = 0.8242

2019/12/18 update

methods

Modified DL framework with fully convs for fast WSI prediction.

We replaced the last GAP and fc. in resnet34 with **AP with kernel size 16x16 followed by fconv and classifer_conv with 1x1 kernel **.

A simple implementation is shown as followed:

# self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
# self.fc = nn.Linear(512 * block.expansion, num_classes)
self.avgpool = nn.AvgPool2d(kernel_size=16, stride=1)
self.Fconv = nn.Conv2d(512 * block.expansion, 512, kernel_size=1, stride=1, bias=False)
self.Fbn = norm_layer(512)
self.final = nn.Conv2d(512, num_classes, kernel_size=1, stride=1, bias=False)

Training predurce is the same as the previous, but WSI prediction is faster than it.

reference1:ScanNet: A Fast and Dense Scanning Framework for Metastastic Breast Cancer Detection from Whole-Slide Image reference2:RMDL: Recalibrated Multi-instance Deep Learning for Whole Slide Gastric Image Classification

Visualization results

• one visualization result is the WSI prediction. We compared with the previous. The result of the previous is on the left, and new result is on the right. Spent time is shown in the figure.

• Another visualization result is the PCA and TSNE features points cluster visualization. Upper is the PCA method and Down is based on TSNE method.

TODO

• Reorganize the codes/
• convert the slide2patch.m to py script

You can refer to Gatsby2016:Fast-WSI-Prediction for more details.