Classification Accuracies of Malaria Infected Cells Using Deep Convolutional Neural Networks Based on Decompressed Images

Transcription

1 Classification Accuracies of Malaria Infected Cells Using Deep Convolutional Neural Networks Based on Decompressed Images Yuhang Dong, Zhuocheng Jiang, Hongda Shen, W. David Pan Dept. of Electrical & Computer Engineering, Univ. of Alabama in Huntsville (UAH)

2 Problem Statement Machine Learning for Automated Classification of Malaria Infected Cells Wholeslide Images Dataset of Cell Images for Malaria Infection Compression Methods Simulation Results and Case Study Conclusion

3 Problem Statement In many biomedical applications, images are stored and transmitted in the form of compressed images. However, typical pattern classifiers are trained using original images. There has been little prior study on how lossily decompressed images would impact the classification performance. In a case study of automatic classification of malaria infected cells, we used decompressed cell images as the inputs to deep convolutional neural networks. We evaluated how various lossy image compression methods and varying compression ratios would impact the classification accuracies.

4 Automated Identification of Malaria Infected Cells 214 million malaria cases, causing 438,000 death in 2015 (source: WHO) In order to provide a reliable diagnosis, necessary training and specialized human resource are required. Unfortunately, these resources are far from being adequate and frequently often unavailable in underdeveloped areas where malaria has a marked predominance. Whole slide imaging (WSI), which scans the conventional glass slides in order to produce high-resolution digital slides, is the most recent imaging modality being employed by pathology departments worldwide. WSI images allow for highly-accurate automated identification of malaria infected cells.

5 Machine Learning for Malaria Detection Classification accuracy of feature-based supervised learning methods: 84% (SVM) 83.5% (Naïve Bayes Classifier) 85% (Three-layer Neural Network) Deep learning methods can extract a hierarchical representation of the data from the input, with higher layers Red blood cell samples representing increasingly abstract concepts, which are increasingly invariant to transformations and scales. Study how lossy decompressed images would impact the classification performance, by evaluating LeNet-5 (one CNN) on four methods: Bitplane reduction JPEG and JPEG 2000 Sparse autoencoder

6 Wholeslide Images of Malaria Infection However, there is NO publicly available high-resolution datasets for malaria to train and test deep neural networks Need to build one! Image of with 100X magnification Entire slide with cropped region delineated in green Whole Slide Image for malaria infected red blood cells from UAB

7 Compilation of a Pathologist Curated Dataset Link to the dataset Apply image morphological operations to extract single cells. The dataset was curated by four UAB pathologists. Each cell image was viewed and labeled by at least two experienced pathologists from UAB Medical School (our collaborators). One cell image can only be considered as infected and included in our final dataset if all the reviewers mark it positively, whereas it will be excluded otherwise. The same selection rule also applies to the non-infected cells in our dataset. The dataset consists of 1,034 infected cells and 1,531 non-infected cells.

8 LeNet-5 Batch 64 label accuracy loss malaria (Data) Batch 100 Batch 100 Batch 64 data conv1 (Convolution) kernel size: 5 stride: 1 pad: 0 conv2 20 conv1 (Convolution) kernel size: 5 50 ip1 conv2 (Inner Product) stride: pad: 0 ip1 accuracy (Accuracy) loss (SoftmaxWithLoss) scale (Power) scale pool1 (MAX Pooling) kernel size: 2 stride: 2 pad: 0 pool1 pool2 (MAX Pooling) kernel size: 2 stride: 2 pad: 0 pool2 relu1 (ReLU) ip2 (Inner Product) 0 ip2 Flowchart of automated malaria detection. 1. Convolutional layer 2. Pooling layer 3. ReLU layer 4. Fully connected layer 5. Loss layer

9 LeNet-5 INPUT C1: feature maps 56 S2: f. maps 28 C3: f. maps 24 S4: f. maps 12 C5: layer 500 F6: layer 500 OUTPUT 2 Gaussian connections Full connection Full connection Convolutions Subsampling Convolutions Subsampling LeNet-5 Convolutional Neural Network Architecture.

10 Bitplane Images The image in the top row is the original malaria infected cell. Each of the four rows has eight bitplane images, with the leftmost column representing the least significant bitplane (LSB) and the rightmost column the most significant bitplane (MSB). The second to forth row are bitplane images from R, G and B channels, respectively. The bottom row are bitplanes for combined RGB channels. In this particular example, the MSB retain less features (e.g., the characteristic ring form of the parasite in an infected cell) than the second MSB, due to a majority of pixels in the original image have values above 128.

11 JPEG and JPEG 2000 Decompressed Image JPEG JPEG-2K The number above an image is the corresponding compression ratio. The higher the compression ratio, the lower the quality of the reconstructed image. For JPEG, The ring form of the parasite is barely visible in the rightmost image, which has the largest compression ratio. For JPEG-2000, The rightmost image has a compression ratio that more than doubles that of co-located JPEG reconstructed image. Reconstructed image still retains the salient features of the original image such as the nucleus and ring form of the parasite.

12 Sparse Autoencoder Artificial neural network Unsupervised learning One encoder and one decoder Use 10 neurons Training process: modify weight and bias to seek minimum difference between input and reconstructed data. Loss in reconstructed data is inevitable, can be treated as lossy compression method. Compression Ratio = # of Pixels 8 # of Neurons 10 8: number of bits per pixel 10: each neuron is a real number between 0 and 1, so 10 bits is enough to represent number from 0 to

13 Simulation Results Classification accuracies of reconstructed bitplane images. Higher bitplane leads to higher accuracy: flipped LSB only changes intensity by 1(2 0 ); flipped MSB changes intensity by 128(2 7 ) Bitplane from combined RGB channels offers higher accuracy, albeit the cost of lower compression ratio. RGB component on higher bits differ a lot leading to different accuracy. 1: keep only the least significant bitplane 8: keep only the most significant bitplane

14 Simulation Result and Discussions The higher the compression ratio, the lower the visual quality, the lower the accuracy of classification. JPEG: higher than 95% when ratio is 10 JPEG-2K: maintain 95% even with ratio of 30. Large reduction of image size will be beneficial to storage and transmission The classification accuracies of the reconstructed images using the JPEG and JPEG 2000 methods

15 MNIST Dataset Left to right: combining the first 8 bitplanes (from MSB), goes down to combining the first 1 bitplane (only MSB). Compression from 1:1 to 8:1. Top to bottom: Bitplane reduction, JPEG, JPEG 2000.

16 Simulation Result and Discussions Classification accuracies of the handwritten digits reconstructed from the compressed images in the MNIST datasets All accuracies are above 95%! Reasons: High contrast between digits and background Bitplane for background are 1, foreground are 0. Even drop all lower seven bitplanes, digits are still recognizable. certain degree of compression artifacts might indeed help the distinguishing features stand out better through deep learning

17 Simulation Result and Discussions Number of Neurons Compression Ratio Compression ratios VS number of neurons in single layer sparse autoencoder The higher the compression ratio, the lower the visual quality, the lower the accuracy of classification. Even when compression higher than 60, classifier can still achieve over 85% accuracy. Autoencoder offers much wider range than other three method, while maintain reasonable good accuracy.

18 Conclusions Large reduction of medical image size would be very beneficial to many telemedicine applications. we compared four compression methods: lossy compression via bitplane reduction, JPEG and JPEG 2000, and sparse autoencoders. Bitplane reduction method had lower accuracy than JPEG and JPEG 2000 methods. Autoencoders were capable of providing a much more scalable compression ratios than the other three lossy compression methods, while maintaining a reasonably good classification accuracy. Further work: improve image quality of more natural images like malaria cell samples using stacked autoencoders.