(370a) Computer Vision Improvised Cell Detection for Lab-on-a-Chip Diagnostics Point Care Solutions | AIChE

(370a) Computer Vision Improvised Cell Detection for Lab-on-a-Chip Diagnostics Point Care Solutions


Srinivas, S. - Presenter, TCS Research
Runkana, V., TCS Research
Lab-on-a-chip point of care diagnostics are the need of the hour for detection of the infectious diseases. Traditional cell counting involve miniature technologies and scaling up of the complex Micro-Electronic Micro-Structure(MEMS) devices would be of a great challenge. Here, we present an effective technology based on advances in Artificial Intelligence in particular to Computer Vision for Blood Cell Detection. With rapid advances in Convolution Neural Networks (CNN), Red Blood Cells (RBCs), White Blood Cells (WBCs), and Platelets detection thereby location identification can be achieved with greater mean average precision accuracy. Thus, paving way for classification of multiple objects inside the image, by drawing bounding boxes around them, so to augment cell counting.

Classifying objects would be computationally expensive in a snapshot utilizing Convolution Neural Networks(CNN). Since the objects of intrigued might have diverse spatial areas in the snapshot and diverse perspective proportions comes about in selecting a tremendous number of locales of intrigued. Regionbased Convolutional Network (R-CNN) & Fast Region-based Convolution Network (Fast R-CNN) are published by Ross Girshick, Microsoft Research in 2014 & 2015 respectively. R-CNN utilizes selective search algorithm for the region search and tends to extract region proposals. It performs the Support Vector Machine based classification algorithm to classify the presence of the object among that candidate region proposal. Fast R-CNN is proposed to reduce the time consumption associated with the high number of models needed to analyze all proposals for the region. CNN Feature Extractor generates a convolutional feature map that is later used to identify the proposal region. Shaoqing Ren, Microsoft Research, 2015 published an extension to the Fast R-CNN, Faster Region-based Convolutional Network (Faster R-CNN). Faster R-CNN introduced the Regional Proposal Network (RPN) to generate proposals for regions, predict boundaries and augments object detection. Faster R-CNN is composed of the RPN and the Fast R-CNN. Thus, Fast and Faster R-CNN methods detect proposals for regions and detect an object in each region. Region-based Fully Convolutional Network (R-FCN) published by Jifeng Dai, Microsoft Research, 2016, involves convolution layers that allow end-to-end back-propagation for weights update during training and inference. It simultaneously takes into account the object detection (location invariant) as well as their position (location variant) & merged the two basic steps in a single model. The previous algorithms for object detection use regions to locate the object in the image. You Look Only Once (YOLO) published by Santosh Divvala & Ross Girshick, Allen Institute for AI & Facebook AI Research respectively in 2016. The YOLO network will look at the elements of the snapshot that have high chances of containing the object. It forecasts bounding boxes and class probabilities with one network during a single analysis. Simplicity of YOLO algorithm permits real-time predictions.

Dumitru Erhan, Google Brain, 2016, published Single-Shot Detector (SSD). The developed algorithm namely, SSD simultaneously forecasts the bounding boxes and then the class probabilities with a CNN architecture from end to end. It eliminates the need for a network of regional proposals. A few modifications, including multi-scale features and default boxes is applied to SSD to recover the drop in accuracy (the accuracy is measured as the mean average precision (mAP), the accuracy of the predictions). These improvements allow SSD to match the Faster R-CNN’s accuracy using lower resolution images which further accelerates the speed. Mask Regional Convolution Network (Mask R-CNN), published by Kaiming He, Facebook AI Research, 2018, adds a parallel branch to the detection of the bounding box by extending Faster R-CNN to predict the mask of the object. An object’s mask is its pixel segmentation in an image. The segmentation of the image thereby groups pixels of the same object. RetinaNet (Focal loss for Dense object detection) published by Tsung-Yi Lin, Facebook Research, 2017, is a one-stage object detector (such as SSD and YOLO) with a two-stage detector performance (such as Faster-RCNN). They proposed a new loss function called classification focal loss, which increased the accuracy significantly. RetinaNet is essentially a feature pyramid network with a loss of cross-entropy replaced by a Focal loss. The results obtained by implementing the above mentioned state of the art object detection algorithms are compared to the traditional baseline of cell counting technologies. Greater insights of the detection architectures & the results obtained would be discussed.