DNN Model Compression

Medical applications in computer-aided diagnosis require algorithms that can often deliver high performance in addition to good accuracy. In addition, there is a growing demand for image processing applications on low-end hardware or embedded solutions.

Solutions using the latest AI algorithms based on neural networks can provide very accurate results for many applications, but the incredibly large set of parameters can get in the way of the aforementioned applications.

That is why at Chimaera we have developed an algorithm that intelligently reduces the number of parameters of a neural network, while trying to maintain the highest possible accuracy.

Data

We have trained and tested our algorithms with a dataset of 67 images of a seven-class segmentation problem. The target organs are: right and left lung, liver, spleen, right and left kidney. This data set was also used for a two-class-segmentation problem (i.e. differentiating tissue from background). As basis for our compressed algorithm we used the popular SegNet model.

Results

 

Number of
Parameters      

Accuracy     

Compression

SegNet 7-Class   

29,458,211

0.930

 

Compressed

7,881,722

0.927

73.2 %

    

SegNet 2-Class   

29,457,552

0.940

 

Compressed

1,213,325

0.929

95.9 %

A typical segmentation result of the 7-class model shown for a sagittal slice on an evaluation 3D CT image (left) together with the segmentation of the organ classes (right).