NYCU MESpeaker: Wang, Hsiang-Sheng, M.D., Department of Anatomical Pathology
Linkou Chang Gung Memorial Hospital
Title： Automatic Mycobacterium tuberculosis detection using simple image processing with artificial intelligence
Date: 2019/11/5 (Thu) 3:30 pm - 5:20 pm
Location: Engineering 5 Building B1 International Conference Hall
Summary:
Tuberculosis (TB) infection remains a major public health issue in Taiwan. In clinical practice, acid-fast stain (AFS) is a special stain for identifying the Mycobacteria. However, it is a time-consuming and exhausting work even by experienced pathologists. Due to the specific feature of TB in AFS, we generate an image filter processing, then combined with convolutional neural network (CNN) to detect TB automatically in virtual slides.
We first create a six-filtered method for image processing and pick up candidates by color, size, shape, color saturation, background correction and edge. The candidate image is cropped into 40 x 40 pixel small images and piping into CNN for recognition. We use tensorflow and keras as our CNN backend. The CNN structure is generated by multiple paired layers of conv2d network and maxpooling layer, followed by 2 dense layers. The training set contains 52 positive samples where TB are labeled by experienced pathologists. Another 50 samples (All TB PCR positive but only 22 cases of them are detected by pathologists with AFS in final report) are also collected for validation.
The 50 validation cases are 22 AFS positive in pathologist’s final report. Our AI takes total 1~27 minutes to go through each case with an average of around 10 minutes. AI finally picks up 47 cases positive within the 50 validation cases and returns the locations of AFS positive areas in each slide. We re-evaluate all these areas and only 3 of these 47 cases are false positive. The rest of 44 cases are all true positive. There is no TB positive case missed by AI.
The AI system we build can work pretty well on detecting highly possible candidate within AFS. Our result also shows that AI is able to assist pathologists to recognize TB even in very small amounts of bacilli.
We first create a six-filtered method for image processing and pick up candidates by color, size, shape, color saturation, background correction and edge. The candidate image is cropped into 40 x 40 pixel small images and piping into CNN for recognition. We use tensorflow and keras as our CNN backend. The CNN structure is generated by multiple paired layers of conv2d network and maxpooling layer, followed by 2 dense layers. The training set contains 52 positive samples where TB are labeled by experienced pathologists. Another 50 samples (All TB PCR positive but only 22 cases of them are detected by pathologists with AFS in final report) are also collected for validation.
The 50 validation cases are 22 AFS positive in pathologist’s final report. Our AI takes total 1~27 minutes to go through each case with an average of around 10 minutes. AI finally picks up 47 cases positive within the 50 validation cases and returns the locations of AFS positive areas in each slide. We re-evaluate all these areas and only 3 of these 47 cases are false positive. The rest of 44 cases are all true positive. There is no TB positive case missed by AI.
The AI system we build can work pretty well on detecting highly possible candidate within AFS. Our result also shows that AI is able to assist pathologists to recognize TB even in very small amounts of bacilli.