Mitosis ICRP2012

Necessary libraries:

• Numpy
• PIL
• matplotlib
• torch
• skimage
• pandas

Flow chart

Input image → Normalization → sliding window crop (512,512) → segmentation model → segmentation extraction (120,120)→ classification model → final results

Training hyper-params

Segnmentation

Adam, 1e-4 lr, no lr decay, 20k epochs, normalized loss + global maximum loss

Classification

Adam, 3e-4 lr, no lr decay, 5k epochs

Pre-processing

Images are normlizaed using the following method
. Macenko M, Niethammer M, Marron JS, Borland D, Woosley JT, Guan X, Schmitt C, Thomas NE (2009) A method for normalizing histology slides for quantitative analysis. In: IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009. ISBI’09, pp 1107–1110. IEEE

In the evaluation folder, original images end with this formate'A00_00.bmp', Normlized images end with this formate 'A00_00_n.png'

Augmentation

• HorizontalFlip(p=0.5)
• VerticalFlip(p=0.5)
• RandomRotate90(p=0.5)
• Downscale_Upscale(interpolation=cv2.INTER_CUBIC, p=0.5)

Los Functions

Segmentation : loss = BCE pixelwise(X,y) + BCE AVG(Sigmoid(X),labels) classification : loss = BCE(outs,labels)

Evaluation

TP: 188 | FP: 29 | FN: 36 | F: 0.8526 | R: 0.8392857142857143 | P: 0.8663594470046083

Classifier model

Classifier model